Ethorobotics of Dolphinoids

Trekking the Pyrobenthic Canyons via Microsubs

Notes of Douglas Moreman, ©1993-95

8154 Rainbow Drive Baton Rouge LA 70809 (225)928-9634

Part I. Dolphinoids for RidgeTrek   

An Exercise in Ethorobotics
Intelligence Serves Desire
Advantages of Autonomy
Microsubs and MotherShip
Conditions in the Benthic Canyons
RidgeTrek: Survey along the Great Thermal Crack
Surviving the Pressure
Aids to Navigation
Mobile NavNet
Stages of MissionOne
Descent; Initialization and Calibration
High, Fast Survey
Low, Slow Study
Dead Reckoning into OuterDarkness
Nodes of a Modular Brain
Memory: a Structure for the Short Term
Maps: Raw Data to Umwelt to Virtual Reality
3D Models of Various Quantities
Parametric Wireframes
Miscellaneous Details on Wireframing
Using Maps
Virtual Trek and VideoSubs
Wireframe Maps Aren't Good Enough

Eigenforms and Umwelt
Behaviors Are Motivated
Curiosity: Interests of a Microsub
Boredom Is Required
Quantities & Spots of Interest
Features(Eigenforms) of Interest
Behaviors that Result from Curiosity
Driven Behaviors
Elements of Ethology for Driven Behaviors
Action Specific Motivations
Driven Chains and Taxis Drives
Deciding What to do Next
Phases of Life
Social Behaviors
Advantages and Difficulties
Social Behaviors and Urges
Language Emerges from Shared Eigenforms
UberUmwelt, SocialUmwelt, and EigenRoles
Structure of Chatting
Calibrations and Play
Sensors; Behaviors
TurnTable (Introducing Kinesthetic Learning)
Other Control Tables
Future Topics


Part I. Dolphinoids for RidgeTrek

The wonders of the great canyon of the Mid Atlantic Ridge lay unimagined beneath the keels of ten thousand passing ships even as the United States landed men on the Moon. Now we know, roughly, the tectonically spreading nature of the canyon, its width, dramatic cliffs and pits, its ten thousand mile length, and fragments of its alien life. We have outlined its mystery. The Ridge exceeds in size and potential relevance anything seen on the Moon. Yet we know the dead backside of the Moon in greater detail. Confronted by this challenge, our government has yawned, but some of us have been inspired. Why not, we wondered, design a microsubmarine around a laptop computer, equip it with sensors and a massive battery

and send it, with pod-mates, on an autonomous mission lasting for a week or more? Our wondering evolved into problem identification and solving. We have made advances in software, as you will see here, and, fortunately, found like-inspired ocean engineers hard at work on the formidible physical problems. Our confidence is growing. With our laptop computers and a spirit of adventurous exploration and enterprise we can be among the first to design an approach to the challenge. We may, at the same time, help launch an autonomous robotics that will reach into factories and homes and the InterNet, changing lives. We begin our part with a small exercise in imagination.

An Exercise in Ethorobotics

A surface ship places and services several unmanned, seafloor-based submarine systems. Each system has one Homebase, usually immobile on the seafloor, and at least one microsub, an un-tethered submarine run by a computer-based brain. Each microsub explores the seafloor, gathering data and, while recharging its power supply, reports information to HomeBase.

Our design exercise is to program the behavior of one microsub. We desire an adaptive system whose exploration will be guided by discoveries made in its past exploration and by changes, possibly accidental, in its own performance parameters. We will attempt to use lessons learned from ethology and other studies of the behavior of animals.

We seek, through this exercise, general principles for the design of behavior of autonomous robots and for the foundations of ethorobotics. Along the way we may, necessarily, pose and solve puzzles about the natural behavior of ourselves and other animals.

We will begin to design the behavior of one microsub. Later we will examine social interactions of HomeBase and a number of microsubs.

In this study you may wish to, as I do from time to time, outline in increasing detail a computer program to implement some part of our behavioral design. This will force the nonsense out of our theories, serving one of the main purposes of more traditional forms of "empirical testing".

Intelligence Serves Desires

To design an "intelligent being", we begin with drives and motivations. The popular idea in pseudo-science fiction of an "android" lacking "feelings" may be nonsense - if feelings are reflections of motivations. Motivations form the basis of adaptive behavior. We represent motivation in a mathematical structure of interacting drives. Adaptive "instinctive" behavior will lie on the foundation provided by that structure. One form of "programming" for a mission will consist of adjusting relative influences of drives, particularly those of "Curiosity". Later, intelligence will grow as some of the instinctive behaviors become ever more flexible: as the "stimulus space" becomes more complex and variable, and as variable and complex habits are shaped by trial and result.


Advantages of Autonomy

One service ship can have, spread over scores of miles, several HomeBases and their microsubs operating simultaneously. A large number of subs can be operated by one team of humans. Microsubs, not needing to periodically surface for the relief of a human crew, can achieve more continuous productive time than can a small, manned sub and will make a rich computer model of the world it explores. There are hundreds of thousands of linear miles of interesting seafloor to traverse and explore to some distance on either side. Microsubs will not be subject to mistakes and inefficiency born of fatigue or boredom. The "resolution", or density of sampling, of a microsub survey will exceed that possible from a surface ship and, in some respects, be less than that from a manned submersible. We will work to insure that, for middle-resolution surveys of deep seafloors, microsubs will be so cost-effective they will open "a new frontier" for science and industry. Looking far ahead we might see that Data, the first intelligent machine, may, rather than an "android", be a dolphinoid.

Microsubs and MotherShip

HomeBase rests on the sea floor. It is a recharging station, possibly a massive "box" of batteries, but also houses a Brain (one or a network of microprocessors), communications gear, a few sensors, and navigational aids. It communicates, possibly at long intervals, with ServiceShip on the sea's surface and, much more frequently, with each of its microsubs. Upon appropriate signal from ServiceShip, HomeBase recalls and retrieves every microsub; then, by decreasing its density, floats to the surface.

Later we will consider HomeBase to be a MotherShip that moves slowly or intermittently.

HomeBase is a communications hub linking, from time to time, ServiceShip with the microsubs and also linking the microsubs. For example, HomeBase can combine and then share back, via maps (See Maps, Page I.6), the data from several microsubs (Social Behavior, Page I.13). Sub and HomeBase "talk" via audio means and visual.

Instrumentation exists for aiding navigation and calibration by each microsub and includes a sound-utilizing system of beacons (autonomous or transponding) by which a microsub can know its position relative to HomeBase. HomeBase also has a sensor to measure water current, a thermometer and various docking signals.

We envision a stay of at least two weeks on the sea floor with the only pauses in research being periods of restoring a microsub's energy.

Each microsub is controlled by its Brain, a system of microprocessors and software, is battery powered, and has navigation and science sensors. Its density is that of sea water. The sub holds sensors for monitoring of the performance for its various subsystems. Sub can cruise for ten hours at 5 ft/sec (about 30 miles). Sub recharges at and downloads data to HomeBase.

The sub's behavior is determined by its ethorobotic program's reactions to input from sensors.

HomeBase will be called MotherShip if it can move along the seafloor to a new resting spot. A group of interacting dolphinoids will be called a pod.


Conditions in the Benthic Canyons:
Real World Anchor Points for Our Imagination

Water absorbs light. Below 1000 feet, humans report that, even at noon, the sea is "pitch black". The range of IR, infra red, sensors is particularly limited because water absorbs infra red frequencies more than bluer frequencies.

Except near geothermal activity, the deep ("benthic") seafloors are cold, about 2C nearly everywhere. The exceptions include spots along pyrobenthic ridges where the sea floor is spreading in two opposite directions -- as between two continents that are drifting apart. Along the crest of a pyrobenthic ridge runs a canyon. In this canyon, the seafloor on one side may be "drifting" west and on the other, east. Stretched rock in the middle occasionally splits, snapping or easing open a crevasse, up which lava may rise and along which lava may spread. Water, in regions along these spreading axes, is periodically exposed, or nearly, to hot magma being pushed upward by whatever great geothermal process has lifted the ridges. In regions near to magma, the temperature of water gushing from one "thermal vent" was found to be in excess of 700F -- hot enough to melt lead, but, due to extreme pressure, the water was not steam.

Precipitating minerals from cooling water gushing upward from hot springs can form "vent chimneys", one giant of which is 150 feet tall.

The bedrock of the seafloor is cooled lava which oozed up on pyrobenthic ridges. Off New Jersey the seafloor basaltic rock is covered in 200 million years of sediment but, as lava, it long ago oozed forth from the same great thermal crack that runs along the crest of today's ridge, part of the mechanism of continental drift.

On the deep seafloors off the continental shelves, living creatures are rare except near the hot water of the pyrobenthic rifts, where they are abundant and bizarre. Hydrogen sulfide in the hot water is food to bacteria, themselves the lowest level in the pyrobenthic food chains around the world (we presume). In the valley or depression between the parallel and mutually separating ridges of the Great Thermal Crack, the seafloor is mostly black, glassy, cooled lava. Where there is life, there is color, often white, in stark contrast to the usual blackness. Towards the continents the floor is the gray of sediments that are older and thicker as they are further away from the Great Crack.

The pressure is about 64*d pounds per ft2 where d is the depth in feet. At 10,000 feet, the pressure exceeds 300 atmospheres. A balloon inflated to the size of a basketball would, slowly lowered to this depth, shrink to the size of a golf ball, and suffer a commensurate loss of buoyancy. A sudden implosion of a glass sphere brought down from the surface could generate a temperature exceeding that of the compression stroke of a diesel engine.

In many of the deep regions so far visited, the speed of the water current is a few centimeters per second. Though slow, such motion can, near seafloor features, produce vortices: so even without tidal effects, the direction of current is not necessarily constant over time. Along some natural funnels in the seafloor the water may be flowing much more rapidly.

The water may be more or less murky. Hot vent water can be black with precipitates that are forming as rapidly as the emerging water is cooling.

Visibility under artificial light is typically about 50 feet or less -- when the light's source is within ten feet of the eyes.

The speed of sound in sea water is a little less than one mile per second and varies with frequency.

The speed of sound in water increases with density, hence with depth, causing sound emitted horizontally from a spot X near the seafloor to curve upwards. The regions into which sound from X can reach, but for terrain barriers, form the interior of a circle whose radius depends on the depth, the frequency, and the height of X above the seafloor. This limits the range of "beacons."

Sound can be refracted by thermal gradients - which are also density gradients. Sound attenuation rises with frequency.

Some cliffs in the great rift canyons are well over 1000 feet high. Some rare, vertical tunnels left by retreating lava are 1000 feet deep.

There exist vent-precipitated, thiobenthic mounds of metallic sulfides piled higher than 100 feet, 600 feet wide, and more than 3000 feet in length. The metals include iron, copper and zinc. Some parts of some of these mounds survive the dissolving effects of ocean chemistry (oxygen converts metal-sulfides to dissolvable metal-sulfates) and are, over eons, increasingly buried in sediments. Such mineral deposits may exist right up to the continental slopes, and lie waiting. Some, geologically thrust ashore in continental collisions, have been mined for copper. The copper in a mound as large as described above might be worth more than a billion dollars (but try getting to the surface and purified for that price!).


Survey along the Great Thermal Crack

The Mid Atlantic Ridge zigzags for about 10,000 miles from Iceland to South Africa. Along its crest, 10,000 feet below the sea, runs a steeply terraced canyon that is 20 miles wide. The canyon is mostly unexplored and less well mapped than smaller structures on the back side of the moon. The best sonar maps of it, made from the sea surface, do not show features smaller than a football field. It has giant peak-and-pit formations spaced at intervals of a hundred or so miles.

Spaced along the floor of the canyon are clusters of hot springs. Some of these vents are called "smokers" because they remind us of old-time industrial chimneys belching black smoke. Their "smoke", created in the shock of hot water on cold, consists of precipitates of metallic sulfides which slowly sink to the seafloor. Metallic sulfides form mineral deposits.

Hydrogen sulfide in the hot water feeds the bacterial base of an alien food chain.

The "transverse faults", great cracks that run away at right angles from the canyon, may prove to be geologic windows, as yet un-viewed by Man, into the most recent 200,000,000 years of evolution of the Atlantic Ocean.

Today's tools (largely the turtle-speed, 3-man submersible ALVIN) employed at full capacity would require 1000 years to survey just the canyon. Less than 1% of it has been visited so far.

RidgeTrek's survey by a fleet of relatively cheap and swift microsubs will prepare the way for efficient follow-up by human eyes and minds.

MissionOne will test the first microsub on a pyrobenthic ridge in preparation for RidgeTrek.

The sub will seek out regions of geothermal activity along a pyrobenthic canyon and on or near them study:
topography (cliffs, crevasses, chimneys, mounds, and calderas),
flow of hot water (heat flux),
chemicals (including salt),
biology (biodensity, macro and micro photographs),
Sub should detect all sizeable thermal vent-fields and smokers in the search area. Sub will be interested in, yet fear, extremely hot water. Sub should detect all sizeable mounds of metallic sulfides.

Parallel to the mission's main efforts can be others -- such as detection of boulders dropped by ancient icebergs and streaks of sediment snowed down by ancient sea currents near the seas surface. These clues may tell us about ancient climates, hence of possibilities for future climates.

A hundred times as much of our national wealth has gone into space-related efforts as exploration of the seas. Without abandoning our dreams of space-exploration, indeed, in preparation for space-robotics, and at less than a tenth the cost of one year's effort in space, we can send a benthic fleet along the Ridge. The ethorobotics we will develop may later aid the exploration of Mars and other planetary surfaces, atmospheres (Jupiter) and seas (Titan?); as well as serve commercial, Earthly functions. Let us begin the Trek here, in our minds, with one microsub.



Surviving the Pressure

Impressively, a hollow glass sphere of 17 inch outside diameter and of less than an inch thickness can withstand the pressure 10,000 feet below the surface of the sea. Paired hemispheres, with lips polished flat enough to form a water-tight seal, are used today to carry electronics to depth. Small holes drilled through the glass allow electrical connections to devices outside the sphere. Such spheres provide anti-pressure chambers and buoyancy. Another way to provide an anti-pressure chamber is to bore an axial container-hole into one end of an otherwise solid cylinder of aluminum, and cap the opening.

A molded plastic, streamlined hull supports and encloses the spheres and other containers and thrusters. This hull, between pieces of equipment, is filled with water. If our microsub has three such spheres then it might have, including the water in the hull, a mass of about 3/4 ton for the thruster to accelerate. The length would be between eight and nine feet.

Electric motors, appropriately designed, can be filled with oil and yet run. The oil, at great depth, distributes the immense water pressure onto the strong, metal components. Even the electronics for controlling a motor can function when immersed in oil that transmits the deep sea pressure evenly. So, our motors will not need anti-pressure chambers.

There will be a snout, a rod extending forward. This snout will have touch sensors and be used for probing to measure hardness, and can provide some protection in case of a collision. Thermometers on the probe will allow measurements in vents that might, otherwise, be discovered too late to be too dangerously hot for the plastic hull.

There may be, inside the front end of the sub, a small thruster that can, through a tube with an openings on the port and starboard sides, force water either way and so push the sub's bow left or right -- a "bow thruster."

Aids to Navigation

HomeBase has various tools to help

NavNet will be a net of transponders centered on HomeBase and possibly spread over several miles. But, for now, we assume that all navigational aids not on the microsubs are within a few hundred feet of HomeBase. We assume that within five miles of HomeBase, except in areas of "shadow" or echo, our sound-based systems give to each microsub an accurate enough estimate of its position.

HomeBase is the center of NavNet and also has attached to it, for healthy redundancy and for use by microsubs within a mile, its own devices.

HomeBeacon emits sounds whereby Sub can determine direction to Home. HomeRanger acoustically gives information by which Sub computes distance Home.

ShortRangeBeacon is accurate but has a short range. It is based on a bright light or on multiple, high frequency sound-makers and is used for calibrations.

Reflectors of light and sound serve as distinct targets for testing a microsub's ranging devices.

MicroSub has its own tools for navigation

There is a clock and an electronic Compass. Speedometer measures speed with respect to water, not with respect to the ground.

DepthGauge is a accurate to within inches at 10,000 feet. By using a programmed knowledge of tides, Sub computes its Depth below mean sea level.

SonarSounder, a downward-looking sonar determines Height above sea floor. The elevation of the ground below the sub can be computed from Subs Height and Subs Elevation.

A LaserSounder not only can be more accurate within its shorter range, it might also detect those plumes of mineral-laden water emanating from hot springs. An apparent difficulty is that the seafloor on the ridges is mostly black, and so are some plumes.

If plumes give a detectable reflection to low power laser light then our microsub will have plume-detecting lasers pointed in various directions.

ForwardSonar provides terrain-anticipation and collision avoidance. Doppler sonar, if its cost/usefulness ratio warrants its inclusion, can measure velocity with respect to the seafloor. AnticipatorSonars look in directions of possible motion to detect obstacles at some angle to the sides and up and down.

A forward-looking LaserRanger is used within about fifty feet and can check the calibration of ForwardSonar.

Velocity (in its 3 dimensions) can be known better the more money we spend on a velocimeter (giving direction as well as speed). Near enough to the seafloor, a doppler sonar velocimeter is likely to be more accurate than using a combination of speed-in-water, compass-heading, and guessed-at currents.

Attitude can be sensed by means of: a plumb, water pressure differences, behavior of a mass that is spinning or vibrating, or a bubble in a tube of liquid. There is the difficulty of accurately knowing attitude during accelerations, particularly during turns; which argues for some of these means over others.

There is a tradeoff in sonar between either of range and resolution on the one hand and drain of Energy (and also expense) on the other. So, for longer travel between re-chargings, we use shorter range sonar.



Thermometer for detecting changes as small as 0.001C in the water passing by.

Clarity meters measure murkiness of water at Sub and between Sub and seafloor. Murkiness may control the height of Sub, when a view (emitted IR or reflected flash) of seafloor is critical to some part of the mission. Thermal plumes can be found by their murkiness, probably detectable much further from the source vent than is the rapidly falling temperature.

Monitors (feedback) of performance for some mechanical devices (e.g. propeller shaft).

Sensors for the Energy remaining in the battery and the power drain from the battery.

Reflectivity of seafloor both to light and to sound. Measured in different frequencies, reflectivity produces the phenomenon of "color", both for light and for sound. In neither case is this straight forward, range and attenuation having to be considered.

Detectors of life. A simple one might detect white spots on the black lava of the seafloor, as white objects on the ridge come from life. The usual sonars may not be much good at detecting life.

Nose (e.g. a laser spectrometer?). To measurer Ph and of various chemicals (e.g. hydrogen sulfide and salt).

Ears. These microphones compare amplitudes and phase shifts (in each pair of ears) so Sub can determine the direction of the source as well as react to features in sounds, suing sounds as do owls and bats. The most important sounds include those of the navigational net, MotherShip, other microsubs, and a nearby volcanic eruption.

MappingSonars, including the vertically down-pointing SonarSounder and maybe several others sonars pointing downward in off-vertical directions.

Magnetometer measures, locally, the Earth's magnetic field (whose variations relate to past motions of tectonic plates). Sunken ships might be discovered by this device.

Optical sensors allow, by virtue of their higher frequencies, higher resolution of sensing than do audio sensors. Passive optical systems will be used near HomeBase and active systems may be used, frugally, in exploration. Passive IR sensors may, at short ranges, detect really hot water.

Sensors must be maintained. This leads to numerous special behaviors (See Calibrations and Play, Page I.16).

Mobile NavNet

If, for NavNet, we use transponders that are anchored (by weights) to the seafloor, how will we easily move our system? Must a manned surface ship retrieve the system, check it out (e.g. for damage during the retrieving), then re-deploy it in a new position a short way along the Canyon? There is a faster and less costly way. We could put our NavNet onto specialized TransponderSubs whose role consists mainly in accurately holding its beacon or transponder in a selectable position relative to MotherShip. This will allow our survey site to step along the Pyrobenthic Canyon, in our RidgeTrek from Iceland to South Africa. MotherShip and, possibly, the TransponderSubs may occasionally rise to the surface for recharging and refitting.


Stages of MissionOne

We are specific about a mission and about details of its "stages" in order to provide a concrete platform for our thoughts. Our goal here is to begin a theory of ethorobotics. The theory will be improved through generalization from lessons learned in specific contexts. Here are stages of MissionOne:

Initialization and Calibration: Sub descends to the seafloor where it surveys the vicinity of HomeBase (within "sight" of, or nearly, the lights of Home), measuring water-current and land-elevation and running tests of its own functioning. ServiceShip tarries above to see that the system is functioning and seems safe. Calibrations and tests of tools will occur in most other stages as well. (See Calibrations and Play, Page I.16)

HighFastSurvey (HSOne) is a high-altitude survey within effective range of NavNet. Let us assume that the search path is pre-programmed so that, viewed from above, it looks like a turbine fan (or a daisy missing every other petal). The path of the survey then consists largely of bladeshaving straight edges, called "legs", running away from and then, after an off-setting turn, back towards HomeBase. (For remarks on "grids", rectangular and otherwise, see Maps, Page I.6)

Sub moves briskly along, perhaps varying its altitude in slow response to changes in the elevation of the seafloor below. The Height is "high" as a guard against high speed crashes into some steep outcropping of erratic terrain. During non-stop travel, Quantities are measured and recorded in maps: elevation of seafloor, temperature of the water, and others.

The search pattern of HSOne can be fixed, as is the cocoon-spinning behavior of some caterpillars, or adaptive, as is the web-spinning of some spiders. In some extreme terrain the survey pattern may necessarily be altered; there may be high cliffs marking the boundary of the zone of exploration. We can allow adaptability: sufficiently interesting spots can elicit immediate exploratory behavior. Some topography along the Great Thermal Crack is jumbled and will demand flexibility even in HSOne.

After HighFastSurvey, Sub analyses its Map (See Maps, Page I.6) and computes the existence and the Interest (See Curiosity, Page I.10) of various blade-crossing features such as ridges and plumes.

CloseSlowStudy has two kinds of activities whose details depend upon the preceding survey:

FeatureFollowing involves cross-blade features such as ridges and thermal plumes which can be deduced in outline (using data from HSOne). They can be ranked by a measure of Interest. Then, after he computes an efficient schedule, Sub can go to one end of a long feature and follow along that feature, measuring its details in high resolution: sharp gradients on the seafloor, hill tops, and valley bottoms. Sub can follow a plume against the current to its source.

CloseStudy involves maneuvering for close study of spots and objects of interest, such as a vent chimney. The Interest of a spot X is a number computed from earlier measurements at or above X. Terrain features that are nearly vertical can be sonar-pinged and photographed from the side. Mineral-detecting probes can be pressed against or into objects (if the sub is sufficiently delicately agile).

HighSurveyTwo (Dead Reckoning into Outer Darkness) is a survey beginning at the boundary of the region of HSOne, and probes beyond the range of NavNet into the OuterDarkness. The sub uses prominent features of terrain and its map of water currents for safety in navigation. After HSOne, Sub has a topographic map (including features that run off into the OuterDarkness beyond the signals of NavNet). Such features can be used as guides into the OuterDarkness that also lead back toward Home, increasing the safety of travel into regions where accurate Position is not given by NavNet.

The NavNet, effective over a few miles, is a big target for a sub returning Home from the OuterDarkness. NavNet can fail, but the Map allows safe return Home. If on reaching a point computed to be within hearing of NavNet, and not hearing NavNet, the sub will continue to operate in the mode of DeadReckoning that it uses in the OuterDarkness, and it will match terrain features against the Map. Since the Map includes at least the region of NavNet, the sub has a large target in seeking its way Home even if NavNet has failed.

You can think of other phases of MissionOne yourself.

Done with all the phases of his mission Sub can wait at Home for humans in ServiceShip to retrieve the combined Map of all the dolphinoids and decide what to do next.



Nodes of a Modular Brain

The elements of decision and behavior can be implemented as a system of procedures in the software of the dolphinoid's "brain". Some of these routines will be "loops" that run continuously, or virtually so.

We assume that the Brain will be distributed over a network of programmable microchip-based "nodes" that pass variables and messages to each other. This will enforce a natural and convenient modularization on the design and programming of our software. We will speak of a behavior as being controlled by a node or set of nodes of the brain.

Each node, as a computer, runs a continuous loop and can receive messages (variables, strings, arrays) from other nodes. Implemented on different microprocessors, the looping on one node is, but for controlled passing of messages, independent of the looping on other nodes.

Such isolation of constantly running control loops is an advantage of physical nodes, each having its own microprocessor. Since nodes cost money and add physical complexity and will use energy, probably some nodes will be "virtual", and several may run on the same multitasking computer chip.

Many sensors will have dedicated nodes. There will be a node devoted to the ForwardSonar and one for the SonarSounder and one for each of several other sonars. Actuators too may have dedicated nodes. Each thruster and each fin actuator will have its own node.

The Decider Node decides, by comparing values of drives, what behaviors to turn on next.

The Mapper is a number-crunching node that creates maps from raw data. The system of Maps constitutes, for now, the main, growing intelligence of our dolphinoid and, so, the Mapper node will be comparable to a powerful, battery-powered computer having a hard drive.

Nodes might include: Mapper, EigenformDetector, Decider, MemoryManager, DeadReckoner, HealthMonitor, Thruster, Rudder, Fin, NavPosition, Compass, Communicator, Speedometer, Altimeter, Thermometer, Clarimeter, Sonar(K), Laser(J) (where K and J are appropriate integers).

Nodes communicate with other nodes. Altimeter receives a range from each Sounder (e.g. SonarSounder, LaserSounder) and, from DepthGauge, an estimate of depth. From such information, Altimeter computes Elevation of the terrain below Sub and passes this number to selected other nodes, including Mapper.

In regions of NavShadow, Position is not available from NavNet and the Mapper uses a less accurate estimation of (x,y) location based on "dead reckoning". The DeadReckoner constantly computes DRPosition, an estimate of the position of the sub, using knowledge about the sub's motion since being at some other position. Suppose A is a "known" position of the sub and the sub is moving with constant velocity V over a duration d. The computed DRPosition is A + d*V. The accuracy of this estimation of position depends, practically, on the accuracy of A and of V. Changes in V are measured by "inertial" sensors.

The dead reckoning computations are frequently updated by the Position, when available, from NavNet. The HealthMonitor can record or integrate and eventually, through a drive (Behaviors Are Motivated, Page I.9), respond to differences (desired to be null) between DRPosition from DeadReckoner and Position from NavNet.

Those who design ethorobotic software tools should enable the designer of behavior to think of "nodes" in terms of behaviors and to hide from him the messy details of physical devices, chips and such.

Memory: a Structure for the Short Term

The microsubs sensory memory may be processed by more than one computer, each having RAM chips and hard drive(s). We could think, conventionally, of RAM as short-term memory and hard drive as long. But within RAM and hard drives there are memories of shorter and longer duration. Each form of memory storage may be organized into "rings". Ignoring the differences in storage type for now, we look at the ring structure.

Some readers may here wish to skip to the next section, Maps.

BaseMemory is the main storage area for sensory data. It has a pointer PtrSTM to the most recently recorded data item. In the beginning, PtrSTM points to the first position in BaseMemory. Should, in time, PtrSTM reach the "end" of the memory, it begins again at the first position, so that new memories will erase old (except where marked for saving). BaseMemory can be thought of as a "ring" and it can be extended over several physical devices (e.g. gigabyte hard drives).

STM, short term memory, includes (consists of?) PtrSTM and a system of pointers to whole (not-broken-into-parts) pieces of BaseMemory. "Removal" of a piece of STM ("forgetting"?) merely amounts to re-assigning its pointer to a newly arrived piece of memory. [This contradicts the usual view of how "short term memory" operates in humans.]

A pointer is "freed" for reassignment when 1) its piece of memory has been analyzed [and possibly incorporated into a model, see Maps below] or 2) when the maximum number of STM-pointers is in use and another is needed for newly arriving data.

It is possible that some selection and representation takes place between sensation and recording in BaseMemory. Some sensory nodes will do this before transmitting information to the MemoryManager.

Note that PtrSTM is probably a pointer to the STM-pointer that has most recently been assigned a referent.

What, in a dolphinoid, is "long term memory"? There are at least two reasonable candidates for the missing meaning. LTM could refer to 1) the sensory data which formerly was in STM and which has not been erased but is no longer pointed to by the STM systems of pointers or 2) maps and other models created by abstraction from sensory data.

Before physical erasure, memories, even those lacking fast pointers, are available for occasional episodes of deep analysis, which may, for example, be done during recharging.


Maps The Umwelt of the Dolphinoid
Becomes Our Virtual Reality

The main product of our dolphinoid's survey will be a kind of map representing all acquired facts about the terrain. The structure and use of this Map will save our dolphinoid from drowning in a torrent of raw facts, steadily increase his efficiency of exploration, and help him avoid natural dangers in the terrain.

The wireframe models, made famous in applications of computer graphics to cinema, are the basis of presentations to humans and of our dolphinoid's on-going search for meaning. Meaning that is relevant to the world (Umwelt) of our dolphinoid is connected to maps as eigenforms, innate concepts provided by human programmers (or trainers of neural nets). Eigenforms serve as atoms in compressed representations. Eigenforms, common to the members of the exploring team, underlie the atoms of a language for rapid, prioritized communication.

3D Maps of Various Quantities

The sub will have its first map installed before launch. This low resolution map of terrain elevation can be created by sonar from surface ships.

Data of kinds in addition to elevation will be organized by the Brain into its system of maps: its "mental image" or model of its physical world. Maps will be three-dimensional computer models of topography overlaid with other data. Having retrieved it topside, the Scientist can view the map in various ways on a high-resolution computer monitor, showing, if desired, smooth graphics renderings, color-coded, in perspective from a selectable point-of-view. These renderings will, except near some naturally colorful vents, have advantages over photographs, as the real terrain of RidgeTrek hides its details in its nearly uniform blackness. Via software, the scientist can fly his artificial eyes into crevasses, around ridges and can change the coloration and the sources of artificial illumination, bringing out desired details.

The system of wireframe maps, which we will call the Map, is a large component of the dolphinoid's Umwelt, its model of its world. Humans will view the Map eventually, but the dolphinoid will use it first as he refines his mission.

Recorded, or represented, in the Map are measures of some Quantity with respect to (x,y,z,t) or (x,y,h,t) locations, where (x,y) corresponds to longitude and latitude, z is elevation of terrain, h is Height of the microsub, and t is time. Examples of Quantity are:
Temperature of water,
Concentrations of chemicals (e.g. salts, manganese),
Acidity (Ph),
Noise spectrum (distribution of loudness of external sounds over frequencies),
Direction of loud (or identifiable) noise,
Roughness of various measures (e.g. roughness of elevation as indicated in a sonar's echo),
Reflectivity of seafloor and cliff faces to light and to sound,

For most measured quantities Q relevant to maps, there is Rate(Q), the rate at which measurements of Q are made (this is called a "rate of sampling"). The total in-flow rate of map data of all sorts can easily be so high that the available memory for numbers will fill before missions end. In making intelligent decisions, animals necessarily ignore most details of sensory input. Our dolphinoid must transform the raw data of senses into compact, useful forms. We can represent and so compress the data flowing in by using functions defined via a set of numbers called "parameters" of the function.

Galileo Galilei near the year 1600 discovered, through extensive experiments, that the distance d(t) a dense (neglect friction) object falls in the passage t of time can be well represented by the functional formula d(t) = gt² + vt + h, provided the parameters g, v and h are well chosen. That functional representation of d(t) is so economical that few people who know it bother to measure or to learn even one experimentally determined data point relating d(t) to t.

We do not have 3-parameter formulas for representing the various quantities our sub will measure. But, perhaps by using hundreds (probably more) of parameters, our dolphinoid can generate formulas sufficiently accurate for its job.

In this discussion, a "map" is a computational system, not a picture, not just a set of points.

Parametric Wireframes

This section on trading fidelity for speed-of-use may include more detail than you care to study. If you suspect so, skip to the next section.

For each map M, there exists a finite set of "data points" which M represents. Let Data(M) denote that set.

Data(M) resides temporarily (until represented) in short term memory, STM. It, in chunks, may be assigned pointers for fast retrieval (this is rare) or left where it is (see Memory, Page I.6), without special pointers and subject to possible eventual erasure.

By the Error of a map M, denoted by Error(M), is meant the sum of every difference between a data point P of M and M's computed representation of P. Error(M) measures the deviation of M from Data(M). By the fidelity of M, we refer in a more pleasant way, and in opposite sense, to Error(M).

A map M has "parameters" used in the computation of points. Some of the parameters of a map M are data points of M or are near to and derived from data points of M. These parameters constitute a set called Base(M) and are called "base points" of M. Other parameters, called "control points" of M, control the shape of the map in between base points. Each control point has influence in the region of some small subset of Base(M).

For the mathematically skeptical - suppose, in a linear space (it does not matter the dimensionality), that A and B are two points and we wish to define an arc from A to B. We can select a point C and define the function P by

P(t) = (1-t)²(A-C) + t²(B-C) + C

for every number t. As t goes from 0 to 1, the point P(t) goes from A to B along an arc whose shape is determined by our chosen point C. C lies on both the line that is tangent to the arc at A and the tangent at B. The three points A, B, and C might replace in memory and in communication a much larger set of data points lying near to the arc.

By allowing a larger number of parameters one can increase the fidelity of a map but decrease the speed of the map's later use.

Suppose M is a map and G is a set containing Base(M) and other points distributed over the surface of M. A second map M2 can be generated so that Data(M2) is G and Base(M2) has fewer points than Base(M). The deviation of M2 from Data(M) surely exceeds Error(M), but M2 is faster than M. A still faster map M3, of lower fidelity-to-the-original-base, could be generated, and so on.

As the area of the mapped region increases and as time allows, faster maps are generated. It sometimes may be that for only a part of a map M will a faster map be generated; for example, for a broad, boring area.

Measures of fidelity (or of error) can be associated with parts of maps. Faster maps may be better for larger features. Below some level of fidelity, a map may lose smaller features; a level that is good for large crevasses may totally miss small chimneys. Whether Sub uses a faster map or a map of higher fidelity depends upon the needs of its behavior at the moment.

Using Maps

Feature-extraction algorithms may exist for each level of fidelity for detecting thermal plumes, crevasses, cliffs, and so on.

When analyzed, the past can alter behavior. High temperature of water (higher than the usual 2C) in some region of one leg will be associated with high temperatures or murkiness on a neighboring leg and so on, and the existence of a leg-crossing thermal plume will be conjectured. This conjecture will be testedby further analysis of existing data (currents are relevant to plumes) and, later, by gathering new data.

FeatureFollowing will involve traveling along deduced cross-blade structures, gathering yet more data, and enriching the maps in regions likely to be interesting to humans. CloseStudy will visit interesting spots along efficient paths (deduced from the maps) and gather data at higher resolutions. There is the (Kalmanesque) problem, solution postponed, of how to add new data from STM to maps that resulted from processing earlier, possibly forgotten, data.

If NavNet fails, Sub can use its present Position and its compass, speedometer, and knowledge of currents to head for Home. It can trace a path over known landmarks in the Map which are recognized via "pattern-matching".

Miscellaneous Details on Wireframing

This section can be skipped, and maybe should be at first reading.

The base points of a map should be few in number compared to the map's data points. However, the map can estimate vastly many more points lying between the data points - enough to give a human a sense of smooth continuity, nice for display by computer graphics.

Methods, and surely details, of representation depend upon the Quantity being represented. Temperature measurements can usually be mapped "on the fly", as the sub moves along, because temperature in most zones of RidgeTrek will be, practically, a constant, about 2C. The representation in these zones will require just a few parameters and execute rapidly. In exceptional regions of hot water interspersed with cold, more parameters will be used.

Our data is not gathered on a rectangular grid -- and this would be hard to do if we tried. However, our parametric representation can be used to compute a rectangular grid, if desired, or a triangular grid.

Included in every map and associated with some map locations (x,y) are markers that point to places in memory containing non-map data associated with a location (x,y): digital recordings of sound, digital photographs, and raw data of various sorts.

There are numbers in regions indicating the quality or the sampling rate of the data used to generate the map in that region. Some markers are copied into a faster map when it is created.

Representing will not always be done in order of time of acquisition of the data. It could be so, if the mapping algorithms worked just along the legs of travel. But, if an algorithm works across legs of survey (as well as along them) then mapping at a location (x,y) will not usually proceed until neighbors on all sides of (x,y) have been visited.

Reviewing. Let the word "arc", for a few paragraphs, mean a segment of travel (most "arcs" here are straight, but I was trained in rubber-sheet geometry).

The Map will have surfaces fit parametrically over arcs and will also have some arcs not yet covered by such a surface.

Imagine laying, on a large paper map of the zone of survey, string along the survey pattern's legs and then along the narrow path of the sub as it follows features and visits spots of interest. The width of the string represents the width over which Sub takes measurements of the Quantity being mapped; of the "foot-print" of the SonarSounders, for example. The data gathered from the path represented by the string is of high resolution: the footprints of successive sonar pings along the route may even overlap. Yet the Map represents not only these thin regions but also regions between segments of string, even though no measurements may have come from those regions. The guesswork that fills in between adjacent segments is more accurate the closer together are the segments of travel. Our dolphinoid cannot look closely everywhere (even "side scan sonar" has shadow regions) and we must be resolved to some guesswork in our maps. But our dolphinoid must remember where the data points forming the base of its representations were taken at high resolution and where taken at low, and our later presentations to scientists can be marked appropriately. Aware that we cannot look everywhere with maximal focus, we recognize the importance of intelligence, of the dolphinoid's growing wisdom about where to focus its efforts, when to be curious and when to be bored.

NavNet has shadows -- regions where its sounds cannot be heard or cannot be used (e.g. due to echoes). It is from NavNet that the microsub ordinarily obtains updates of Position. When the sub enters a NavShadow, it switches from using Position to using the less accurate DRPosition in creating its maps. A history of paths leading via DRPosition to portions of the Map is maintained.

For every (x,y) location in raw data, and likewise for some, if not all of the other measures of Quantities, there is the number Confidence(x,y). This is a number from 0 to 1 and indicates, as well as the programmer can arrange, the confidence with which the measured position can be used in computations. Confidence(x,y) will decrease with the distance from location (x,y) to HomeBase, the center of NavNet. Confidences can be copied or translated to maps of lower fidelity. Confidence(TerrainsElevation) will decrease as the microsub turns, and so rolls, and so swings the footprint of its SonarSounder outward from the arc of the turn. Elevation, in a turn, is estimated using a measure of the roll.

Sub will have behavior to increase the accuracy of its estimation of position. In some cases, Sub can exit NavShadow and get a fix on its Position by climbing. Additionally, the maps can be corrected by experience. If Sub finds familiar object X offset from the expected location, the location of X will be reanalyzed and not only X in the map, but objects located by reference to X, will be moved accordingly.

VirtualTrek and VideoSubs
Enlist 10,000 Volunteers

There is an interesting technology to dream about: the sub could map its digitized photos, showing plants and animals, onto a 3D wire frame model. We will be able to "see" vistas that no man's eyes will ever directly see, illuminated by a "sun" that these vistas will never see.

Virtual Reality may be used for teaching and entertainment and as a means of presentation and investigation of different categories of data for research. A real-time link from seafloor to a buoy or ship on the sea surface is not inexpensive, but possible. Given that link we could then use existing communications satellites (NASA) to transfer daily or weekly upgrades of the Map to land. The virtual world of RidgeTrek, via CD ROMs, can be explored by thousands of interested people. These explorers can be children in school or home, or experienced scientists.

Or, the World can look over the shoulders of Science as the exploration progresses. Once a communications link from seafloor to land is made, RidgeTrek's Ethosystem will allow real-time interaction via slow television. Scientists and their students will be able to use VideoSubs to remotely view and WaldoSubs to manipulate, seafloor objects almost as if they were there in the submersible ALVIN. This attractive topic leads us astray, but we do pause to suggest that such real-time participation by humans will not remove the time-saving advantages of a semi-autonomous fleet of autonomous microsubs. Rather, our scouting, mapping fleet of dolphinoids will economize and make more valuable the time of the human operators of video-toting microsubs.and the humans in the deep submersibles such as ALVIN and Deep Rover.

Wireframe Maps Aren't Good Enough

Returning to Earth, as it were, and to the Map, we now must come to understand that the Map, with all its layers of complexity has too few elements of meaning. Intelligence in the behavior of robots will require more direct meaning, in our Map, relevant to behavior. Again, we seek inspiration in that evolutionary proving ground: the behavior of animals in their natural environments.


Eigenforms and Umwelts

The German word Umwelt was used by early ethologists to refer to the world of an animal, as experienced by that animal.

We do not see what animals see - when looking at a flock of gulls or listening to their calls we do not recognize individuals. Using the best of microphones we cannot "see" what a dolphin is describing. We cannot see those patterns of ultra-violet reflection on flowers that lead a bee to nectar. We are not even fully conscious of all details of our own umwelt: few people have consciously noticed eyebrow flashes when making visual contact with a friend, though they have been unconsciously doing it and responding to it every day.

Animals begin life with abilities to recognize important patterns in their sensory input. Consider the following, classic, barnyard demonstration by European ethologists. Over the heads of inexperienced ducklings is sailed along a wire a cardboard silhouette of a hawk: with a short neck in front and a long tail behind. The ducklings do the appropriate innate thing: scurry for cover. It is interesting that when the same model is sailed along the wire backwards it does not cause alarm, perhaps it looks like a long necked goose.

The word eigenform refers to a feature of the world of an animal that is, in some sense, recognized innately, as in the example of a hawk to a duck. A good part of the umwelt, the personal world, of an animal may be composed of instances of eigenforms. Indeed, in ethorobotics, we may succeed in broadening our understanding (or definitions) of eigenforms to include sufficiently many of the fundamental elements of perception that a large "space", the Eigenwelt, lying in the umwelt of a dolphinoid can be defined mathematically in terms of them. The Eigenwelt will model all stimulus releasers of innate behaviors.

Instances of eigenforms will be called eigenthings

For each eigenform, there will be program modules for detecting its eigenthings. Their algorithms will operate both on raw data and on the Map's representations of data.

Eigenforms important to RidgeTrek include those hot springs called "vents", vent chimneys, vent plumes, vent mounds, dead vents, flat-faced cliffs, crevasses, and volcanic cones. One eigenform will represent a largish, unexplored region surrounded by regions already mapped. Such terrain-related eigenthings will be some of the fundamental elements for computations of intelligent behavior, based upon motivations.

Amongst animals, for some eigenforms there can be (human-made examples show) super-normal eigenthings. In ethology these have been called "super-optimal releasers". These human-made distortions of eigenforms have a stronger effect on the animals of study than do more natural versions. This phenomenon may guide us in our future studies of "generalization", of distances in the Eigenwelt stimulus space. Newly popular tools of "neural nets" and "fuzzy logic" are relevant here, but others may be better.

Eigenforms provide a symbolic and conceptual means of increasing the efficiency of a Map. Rather than hundreds of thousands of pixels representing a 3D picture of a particular vent chimney, and rather even than a few thousand points in a parameterized model, our Eigenwelt software will note that in a certain location a chimney exists. It may also record a few measurements such as the height. This compression of information through abstraction may be close to the way animal minds work and greatly increases the speed of computational decision-making and communication.

Photographs, in a process called eigenforming, could be compressed for storage and communication via eigenforms. Consider the dryland case of photos of human faces which could be represented by eigenforms such as "nose" together with a reference (by a software "pointer") to a particular nose-type in a "library" of examples of facial components. Actual brains may use something like a neural net for each component (in robotics, one net of a system could be trained to classify noses).

Behaviors Are Motivated

We presume that natural intelligence rests in part upon willfulness, that natural urges are clever adaptations based on a hundred million years of experience and constitute an evolved wisdom.

What the dolphinoid does, which of a number of mutually exclusive behaviors are turned on, is determined partly by a comparison of the competing motivations.

GoHome is a behavior. Home is where the dolphinoid can replenish its energy, calibrate its sensors and behaviors, chat with other pod-members, and answer a summons.

There is a drive, our program's model of motivation, that is specific to the behavior GoHome. This drive increases as Sub's energy drops towards the estimated energy needed to get home. It also increases with some measures of poor performance (e.g. power-consumption at some speed exceeds that predicted by the SpeedTable, Page I.18).

When the drive to GoHome gets large enough, the dolphinoid stops whatever else it is doing and heads for HomeBase. But, in the presence of features (eigenthings) of high Interest (See Curiosity, Page I.10), the sub may tarry longer before an increasing drive to GoHome overcomes a strong drive to investigate. (You may wonder how this happens since both drives may reach their maximum. See Deciding What To Do Next, Page I.12)

A high value of Interest can increase the drive of a system of investigative behaviors such as that associated with locating the seafloor source of a vent-plume and taking both high resolution measurements and photographs around that source.


Curiosity: Interests of a Microsub

Sub will be "born" (launched) with "innate" interests and will develop, by our rules, new interests through experience. Sub is curious about oddities.

Measurements of high Interest can, sooner or later, influence behavior via the drives in our model of motivation. An Interest will be a number assigned, in many cases, to an eigenthing. There will be behaviors for studying such an instance of an eigenform. The drives, called "curiosities", of these behaviors will increase as the Interest increases.

Boredom Is Required

After our dolphinoid has been studying a region of high Interest, he must eventually move on. There are at least two motivations for quitting: a lowering of Interest and an increase in repulsion. Permanent decreases in Interest of some eigenthing occur as members of a set of investigatory behaviors are completed. A temporary decline in Interest begins after some delay preset by the Mission Programmer who can also preset the amount by which the repulsive drive generalizes (over terrain and to similar regions). Sub can behave as if frustrated, and, as a result, work harder or quit work or even flee (in "disgust"?).

Quantities of Interest

For each Quantity Q there is the function Interest(X,Q) where X takes on values of measurements of Q. Every new measurement M of Q is compared to previous measurements, and to pre-conceived thresholds set for the scientists by the Mission Programmer, and may result in a change in the function. For example, suppose Q is water's temperature. A measurement H of temperature that exceeds all so far may cause Interest(T,Q) for lesser values T of temperature to be smaller than it would have been before; the function changes.

The Scientist may define "oddness" and plan for the microsub to be interested in unusual values of Quantities Q such as:

Elevation (e.g. cliff, tall vent chimney, large mound),

Temperature of water (400C should elicit overwhelming curiosity),


Sonar roughness (varieties of lava on seafloor),

Reflectivity of surface (to light, to sound) at various frequencies,

Concentration of organics, salt, oxygen, hydrogen sulfide, metallic sulfides (some of these may be a "scent" of a distant vent),

Velocity of Current (speed and direction),

Loudness (e.g. of seismic sounds from below or the signature, if such can ever be found, of a smoker).

Or Q may be a gradient-like derivation (steepness) of one of the above. You may have noticed that in the very first example above, "cliff" actually was an example of a high gradient. A long slow rise to the height of the cliff may not be as interesting. "Edges", regions of high gradient, may be interesting for many quantities, including temperature. There are reasons from geology for making-interesting flat cliff-faces, according as they are flat and tall.

Local extremes may also be interesting.

Total Variation Xk - Xk+1 (over some region) might be another Q to make a region interesting.

Surely you can, yourself, invent several interesting measures of Interest.

For some quantities, changes over time will be interesting. Certain ratios may spark interest.

Interest(X,Q) is determined by thresholds set by Scientist before launch and by extremes of X experienced so far. It can be that, for some Quantity, no observed X is interesting (above Scientist's Threshold). Some initially interesting values can become boring because they are dwarfed by experience with more extreme values. The defining parameters of the Interest function for a quantity Q can be altered by such means as a weighted average of recent-most extreme measures of Q.

Spots of Interest

Let us consider how Interest might be computed. Suppose P is a terrain location above which measurements were made or inferred by mapping. Suppose t is a time, and for each Quantity Q, just one measure m(Q,P) exists and I(P,Q) = Interest(m(P,Q),Q) at time t. Assume Scientist has provided priority-coefficients pc(). Assume all Interests have been updated at time t. Then, at time t, Interest(P) = [pc(Q)*I(P,Q)] over all Q.

Let w(X) denote an appropriate weighting inversely related to the distance from X to P:

BroadInterest(P) = Interest(P) + [w(X)*Interest(X)] for spots X in some neighborhood of P.

An efficient interest/duration maximizing path can be plotted connecting some of the spots of interest, when time allows closer study.

Eigenforms of Interest

A feature is a sequence of spots deduced to be (probably) parts of a physical structure of interest. We create algorithms to perceive ridges, thermal plumes and other structures of interest to us. The kinds of features which a microsub M can perceive (and to which it can respond) are some of the eigenforms of M.

Our dolphinoid will be particularly curious about certain eigenforms made innately interesting by the Mission Programmer before launch.

The Mission Programmer must provide weights -- values for calculating the Interest of an eigenthing. A vent should be more interesting the hotter its water, the greater its signs of biological activity, the taller its chimney, the murkier its plume.

For each eigenform, there are specific behaviors: for verifying the existence of an instance E of that eigenform, for close study of E and, perhaps,  for using E for some purpose, such as a waypoint in dead reckoning.

Some of the eigenforms about which our dolphinoid might be curious include vents, plumes, mounds, cliffs and crevasses, flows of hot lava, and the larger unexplored regions that are surrounded by regions already mapped.


Behavior that Results from Curiosity

Curiosity is based on drives that make various investigative behaviors more or less likely. Some of these behaviors can be elicited during one stage of the mission and not another.

Rapidly changing terrain elevations beneath the sub can increase Rate(SonarSounder) and can, if the sub is close enough (20 feet or so) to the terrain, increase the rate of photographing.

A sufficiently large increase in water temperature can interrupt travel and elicit local, vertical gridding, a kind of helixing, through a probable plume.

Likewise, an abrupt drop of terrain below, can lead to increased motivation to do a horizontal grid, creating a high-resolution map of the mouth of a volcanic pit or of the lips of a crevasse.

A place of high Interest may be visited at low height and low speed during a period of CloseStudy. Sub maneuvers in a chosen region (which could be circular and centered on an interesting spot) gathering some kinds of data at higher resolutions and gathering new kinds of data such as photographs, and data from new points of view (sonar, laser, or photo) from the sides as well as from above. Relative values of Curiosity lead to visits to some places rather than to others.

At a vent, behaviors are released which lead to a local, high resolution map of heat flux (temperature and micro-current) and of biodensity.

In addition to studies of places, Curiosity motivates FeatureFollowing. Features, having linear and maybe planar extent, include plumes, mounds, cliffs, and crevasses, whose extent can be traced out. More generally, Curiosity motivates behaviors appropriate to the study of each of several eigenforms.

A plume can be followed up-current to its source.

A thiobenthic mound is followed along most of its extent. At intervals, Sub moves across the mound at right angles to its axis (parallel to the average current), gathering data at high resolution to model its cross-section. Mounds probably can be detected by sonar-softness and studied by special penetrating sonar. Sub might push against a mound to further test its hardness. There may be means to estimate the mineral make-up of the mound; though we intend to avoid tasks better left to other tools such as WaldoSubs and manned submersibles.

A flat cliff face probably results from a geologic up-thrust. The lower parts of a cliff were, long ago, as deep below the seafloor as the cliff is high, and formed one side of a crevasse or a thin crack. On these cliff faces, one might expect to find remnants of hydro thermal action, perhaps the subterranean part of an ancient vent. At those lower depths in the rock, the temperature of the rising water was higher and, where high enough (perhaps only at lower depths below the seas surface), may have caused the precipitation of pure gold. This theory may prove false, but remains of old vent roots will have value to Science. Sub traces a cliff along its horizontal extent (seeking remains of thermal action) with occasional studies of its vertical extent.

We can devise means (e.g. via neural nets) of detecting patterns in sensory data that are predictive of the existence of a vent of particularly high Interest (e.g. one having a high thermal flux or a high biomass). Such patterns can then, upon occurrence, cause a vent to be assigned a high Interest (before the heat flux is measured). This exemplifies "associative learning".

Driven Behaviors

Elements of Ethology for Driven Behavior

An F.A.P., a fixed action pattern, is an innate behavior made more or less likely by a specific drive and stimulus. A reflex, such as an eye blink, is released by a stimulus. A taxis is a reflex that aims or tunes an FAP.

The fly-catching flick of the tongue of a toad may look like a reflex, but seems to depend on a drive (hunger) and to have taxes.

An FAP is an "innate" (somehow programmed during the genetic unfolding called "ontogeny") action that we observers can recognize in the behavior of many individuals of a species. An FAP is said to be motivated (driven) by its S.A.P., specific action potential, its own proper motivation. An FAP is said to be releasedby an appropriate stimulus input called its "releaser".

The stimulus-releasers of innate behaviors of animals are "eigenforms", but have not yet, to my knowledge, been used in zoology as a subset of the basis of a mathematical model of the Eigenwelt of even one animal.

In their own search for inspiration and guidance, zoologists might learn from the new "compression" techniques used in representing photographs of faces. The software seems to use eigenforms such as Nose and a library of noses to reference with just a byte or two. Has Nature, unbeknownst to human inventors, anticipated their "new" methods of compressed representation?

A human FAP, or maybe a reflex, that seems to work on a subconscious level (both in action and in effect), is the rapid rise and fall of eyebrows that precedes a smile as eye contact is made with an arriving friend (who gives the same signal).

FAPs may occur in a chain R1, R2,..., Rn that is just as innate as are its component actions. One action in the chain produces the releaser of the next.

A bee going to a flower to get nectar goes through a chain. It sees a flower and approaches. Near to the flower it switches to a landing mode. Having landed, it switches to walking in search of a likely place for nectar, where it switches to probing for nectar with its proboscis.

What appears as a chain may be based upon a "tree" of innate responses where the branch taken depends upon circumstances (internal or external). A chain may also come from a "funnel" of possibilities.

A "reflex", say some, is like an FAP, but lacks a motivational component and always occurs with about the same intensity. Some reflexes can be made more or less probable by drives that move a stimulus threshold.


Action Specific Motivations

Action Specific Motivations

The behavior of our dolphinoid will be largely determined by "drives".

For each driven behavior B, there is Drive(B), a drive that is specific to B. Each drive D has, at each time, a value from 0 to 1 called DrivesValue(D). If B is a driven behavior then, "SAP(B)" means DrivesValue(Drive(B)) and is read "the S.A.P. of B" or "Drive of B."

For some B's, 1) SAP(B) increases over time [towards its maximum of 1] and 2) SAP(B) is reduced when B is executed. An SAP whose value changes due to passage of time since the last performance of its act will be called a TimeDrive. Using a linear (piecewise) model, the TimeSlope, the rate of change of a TimeDrive may be altered by another drive, or by an event.

We can see in a house cat an example of a drive that feeds another drive. Suppose B is some component of hunting behavior (e.g. stalking, pouncing on, biting the back of the neck of a rat). As Hunger grows, so does SAP(B). However when Hunger is kept low by ever-present cat food, SAP(B) can none-the-less increase until B is released (occurring even in the absence of anything like an appropriate natural stimulus). After B is released, SAP(B) decreases. This concept of SAP (and of Leerlaufreaction) was presented by Nobel Laureate Konrad Lorenz, one of the fathers of Ethology.

A behavior B such that SAP(B) is always 1, is called a "reflex".

For some Bs, SAP(B) changes due to circumstances such as the occurrence of some stimulus. A clear example of this, in animals, is Fear (elicited by some specific stimulus such as sight of a predator). Fear increases the SAP of some Bs and, for others, either decreases SAP(B) or increases Inhibition(B) (not yet discussed). When listening is important, Sub reduces its own noise-producing behaviors. When testing by some other visual means the conjectured existence of a plume, Sub may suppress the strobe light used for photography.

Driven Chains and Taxis Drives

Three examples of driven chains are: tracing the survey pattern of HSOne, tracing a computed feature-following path, and Going Home to Recharge.

Suppose that GoHome is a behavior by which Sub gets to within a distance HomeRange (50 feet is a good first guess) of HomeBase; Dock is a behavior that results in Sub entering a docking bay; Recharge is comprised of some act analogous to plugging in, followed by communicating with HomeBase while monitoring the recharging.

As the energy of Sub's battery drops, a drive, that we might as well call Hunger, increases towards 1. The increase in Hunger causes an increase in SAP(Recharge) and this in turn causes an increase in SAP(Dock) and this causes an increase in SAP(GoHome). Did you notice that the ordering of the increases in the drives is reverse to the occurrence of the behaviors?

Suppose that our dolphinoid is away from Home. When DrivesValue(GoHome) is large enough, relative to competing drives, the GoHome behavior takes over, but neither Docking nor Recharging occurs yet because their releasing conditions are not met. When Sub reaches HomeRange, DrivesValue(GoHome) drops to 0 and the behavior Dock takes over because its drive is high and its releasing conditions are met. And so on. There is more complexity as we shall see.

Some behavior B may have components. Conditions experienced when and after B is elicited may change the drives of the components and so shape B. Such shaping or directing drives are called taxis drives.

Deciding What To Do Next

Whether a behavior B is activated depends on both 1) a measure resulting from the joint effect of SAP(B), the degree of the releasing circumstances, and Inhibition(B), and 2) the competition of that measure with those of competing behaviors.

For making your own examples for the following concepts, you may enjoy thinking of curiosities as systems of drives for investigative behaviors relating to eigenforms.

For each driven behavior B there is Urgency(B) which is computed using:

Urgency(B) = Urge(B) * Conditions(B).

Urge(B) is computed using:

Urge(B) = SAP(B) * Priority(B)

where Priority(B) is a parameter set by the Mission Programmer before the sub is launched. The number Conditions(B) depends on sensed or computed external conditions (the "stimuli" of American psychologists and the "releasers" of European ethologists), on internal conditions, on a computation of practicality (including cost), and on memory (on memories evoked by the other conditions).

The Mission Programmer gives highest Priority to behaviors relevant to survival. GoHome has a much higher priority than SnapPhoto.

The Decider node receives from other nodes, for various behaviors B, the signal Consider(B). Then Decider sorts those behaviors by descending value of Urgency(B). For each of these B's, there is the number EstimatedEnergy(B). Starting at the top of the sorted list, Behaviors B are turned-on if 1) Energy (total safely available) exceeds the sum of EstimatedEnergy(C) where values of C include B and every turned on behavior preceding B in the list and 2) B can be concurrent with all already turned on behaviors. There is a table of Concurrencies created pre-launch and a dynamic table of EstimatedEnergies.

The Decider node runs continuously.

Opportunistic insertions of behaviors are allowed even while some behavior B that should be completed is active. Decider may get a message Consider(C) for some Behavior C that is compatible with B or that would only interrupt B briefly. If Urge(C) is sufficiently high, C will be activated. Examples: a photograph may be taken opportunistically during an OutwardLeg; a vertical or horizontal grid pattern may be elicited during a leg; a Scout may stop to Chat with a neighbor (Social Behavior Page I.13).

Some behaviors are to be avoided. Do not fire a broken strobe light. Do not push against a surface if the nose probe has been broken off. You are left to ponder: "How is a behavior to be cut off?"

We have omitted details, as you may have noticed, partly to spare the overtaxed general reader and partly for proprietary reasons.

Phases of Life

A honey bee, on "birth" from its cell, begins a sequence of phases. For each phase there is a set of driven (or so I presume) behaviors. Some of the phases are, roughly, doing house-cleaning chores, tending the queen, guarding the entrance to the hive, scouting for nectar or for water, gathering nectar or water. The phases are distinct: while doing the functions of one, a bee does not ordinarily perform the functions of another. A bee progresses through the sequence of phases as it ages.

A female ring dove begins a sequence of phases of behavior in early spring. The appropriate ratio of day-length to night-length leads the hypothalamus to send chemical signals down special neural axons to the anterior pituitary, which in turn sends chemical signals via the blood to various reproductory organs that emit hormones. Watching the mating dance of a male ring dove can then lead to other internal changes which lead to motivations of other activities. In appropriate temporal sequence occur the sets of behaviors of mating, nest-building, egg-laying, egg-setting, chick-feeding and so on. Some of the behaviors appropriate to one phase will not be released by innate stimuli appropriate to another phase.

There is some similarity in the behavior of our dolphinoid to "critical periods" in the unfolding of instinctive behavior and to seasonal behavior.

We will here assume that, in MissionOne, Sub is, at each time t, in exactly one of the following primal phases.

Descent. Check for leaks. Check depth gauge. Talk to HomeBase and to humans (checking communications)

Initialization. SystemChecking, Control Table



CloseStudy [with Feature Tracing]


WaitAtHome (HomeBase can order Sub to do this),

Ascent (to Service Ship),


The behaviors that are appropriate to the present phase may be turned ON or OFF by the mechanism, not yet discussed, that sets the TRUE or FALSE values of Consider(B), and that controls some higher level drives that feed other drives.


Social Behavior and Emerging Language

Sciences of animal behavior offer us some lessons in the behavior of groups or of individuals in groups. Darwin's study of the body expressions of dogs has a flavor of fuzzy logic, applied to eigen expressions. There are also the grand lessons from ants and bees. From them we see that small, local, acts of many individuals combine, leading to intricate physical structures or to moving objects too large to be moved by one individual.

Less seems to be known about language in animals. From our work here I think we may come to suspect that such language goes well beyond what has yet been understood. It is not so wild a guess that language emerges in evolution from a basis of fundamental urges and eigenforms.

Advantages and Difficulties

Advantages to a mission of using interacting microsubs include: mutual assistance, speed, and specialization. Also, a mission is more likely to be completed by a pod than by a lone microsub, who could "die".

Two scouting microsubs working from one HomeBase double the data-collection-rate at less than double the cost. And we will see other advantages to having a pod of cooperating microsubs. So let us assume a team of Scouts and also a smaller number of TransponderSubsserving as part of a moveable Navigational Net. We replace our image of a static HomeBase with MotherShip, a submarine, massive with batteries, that moves intermittently along a chain of survey sites, and serves as a navigational and communications nexus for the pod.

Audio communications can fail at longer ranges for larger numbers of microsubs, as the messages will jumble together (we are severely limited in "band-width").

The usual "transponder net" is not appropriate for a pod having numerous scouts; them tending to query the net at about the same time. We will not solve (or clearly state) the problem here, but merely note that its solution can have behavioral, computational, and hardware components.

We must design a mechanism for the resolution of factual conflicts (such as differences in maps of the same region from two scouts). In general, as we design the software, we must be imaginative in our anticipation of things to go wrong. This may deserve even more of our time as does the originally conceived behavior.

Those Readers with appropriate experience may cringe at each suggestion of an interesting behavior for the microsubs; suspecting that a year or more of programming labor might be needed for the implementation of that one behavior. Additionally there are possible interactions of each new behavior with each of the old. We see here a need for tools to ease these implied labors and the beginnings of an argument for building such tools before building many serious autonomous submarines. We will later begin (barely) to show how microsubs can (and probably should) be programmed to program themselves (See TurnTable under Calibrations, Page I.16). We create initial, roughly hewn behaviors which the subs refine through experience with their environment, which we can only roughly anticipate.

Social Behaviors and Urges

Social behaviors necessarily include docking, avoiding collisions, and deciding in case of conflict which sub docks next. More interesting will be exchanges of information, including self-information of social relevance, and division of labor.

Subs will be able to detect other subs by the sounds they make, notably their sonar. Since sonar is highly directional, there will be an omni-directional caller. Scouts will have a drive to CallOut that is higher the more likely another sub is to be present. This reduces collisions and helps to initiate social interactions.

Microsubs will share information relevant to the Map and to Curiosity. The shared Map built by two subs grows twice as fast, and the "wisdom" about where to put their efforts grows twice as fast. Given the overhead of MotherShip and TransponderSubs, a pod will more likely have ten scouts than just two.

A "buddy system", whereby our scouts will travel in pairs (or more), will have advantages in safety and survival. Also, if we allow close study of things of particularly high Interest as soon as they are discovered, a pair can be more efficient than one alone: one can call the other to aid in a rapid study. At first glance, the mathematics of this improvement in efficiency seems to depend upon the value of finishing a high resolution study of an Interesting feature exceeding the value of finishing a low resolution study (survey) of a boring feature (the general seafloor).

Subs are programmed to rescue or to assist in the rescue of other subs. Our dolphinoid can use shared social information to try to locate and report the position of a downed companion (I suspect that in deep water and with humans weeks away, sinking in "death" is better than rising). If a sub loses its navigational abilities or if its ForwardSonar is broken in a crash or burned in a hot vent, the blinded sub will call another, maybe its "buddy", to guide it Home.

People have an urge to download; feeling a compulsion to report on odd events (and on events not so odd). When two Scouts meet and one has recently found eigenthings of high Interest, it has a high drive to download the interesting information to the other sub. As time passes since Sub last interacted with another, its drive to Report increases. The Drive(CallOut) is fed by Drive(Report). Also, a scout may adjust, in some minor, low priority way it's behavior (e.g. path traveled, loudness of CallOut) to increase the odds of crossing paths with another Scout. Such simple alterations of behavior can increase the flow of information in the pod.

Some brief encounters between subs, called chats, demand efficiency, so that the information most likely to be useful is transmitted first. The kinds of analysis we apply to Curiosity (remember, from Page I.10) may be used to compute Interest-to-Another for recent discoveries, recently obtained information from social interactions, health problems, and so on. The ranking of "topics" by Interest-to-Another can order the presentation of topics.

The first, most primitive, chatting will be a downloading of map-parts prioritized by Interest. It will not involve sentences. Each sub, before chatting will subdivide its information to be shared into segments. Each will speak a segment then listen to a segment (for more details see "Structure of Chatting", Page I.15). Each speaks from a prepared script.

Boredom may be worth programming into Chatting, though not so important as for ending a study of terrain.


Language Emerges:

Eigenforms to Icons to Sentences

There will be eigenforms of social interaction, and simple signaling and sharing of data. Of more startling significance will be the probable emergence, upon a base of shared eigenforms used for "thinking", of a common, sentential language.

Eigenforms of uniquely social relevance to each dolphinoid include MotherShip, TransponderSub and Scout.

Upon detecting one, a scout "names" an eigenthing. These names (in a program, they can involve indexing integers) are shared and then used in later communications between Pod members.

Just as eigenforms allow both compressing a database and speedy computing of tasks ("Eigenforms and Umwelts", Page I.9), they also allow speedy communications. A first dolphinoid can tell another of the existence of various named eigenthings at various map-locations and then add details about these things in prioritized order, while time allows.

Using a few "rules of grammar", we can start the construction, or "emergence", of a language of eigenforms and a concept space based upon its sentences (e.g. "Hot vents usually have a high density of life"). The first use of this "language" will not be in talking, but in a kind of thinking, problem-stating and solving, in the "mind" of a lone dolphinoid. The roots of language (with a grammar as well as signifiers) may, in Nature too, be "pre-verbal": used for thinking before used for talking.

The "minds" of animals seem to use eigenforms and animals do convey information about states, internal and external. A crow, it is said, can call so as to say "There is some wonderful food over here where I am" (or, maybe, "Food, Here"). But, we do not know that a crow can use proper names or modifiers to reduce the ambiguity of references: "There are tasty pigeon eggs in that red barn." But, we expect to have our dolphinoids talk to each other using proper names and modifiers, and dolphinoids are not, otherwise, as clever as crows. It has been alleged that some twin human children have developed, or had "emerge", their own language. Might not a language of sorts emerge in a family of crows?

Perhaps before the year 2000 we will see a demonstration of an emerging language of eigenforms in a pod, or simulated pod, of computers. When that occurs, perhaps some zoologists will take a closer look at crows.

UberUmwelt, SocialUmwelt, and EigenRoles

The growing Map, created as a joint enterprise by communicating pod-members will be called the PodsModel or, for more general application, the UberUmwelt.

Deep communication, such as occurs between a scout and MotherShip, is called melding. Information transferred includes the lower levels of abstraction down to some raw data (see Maps, Page I.6).

It might seem likely that MotherShip (and maybe the scout most recently recharged) will have the most complete version of the UberUmwelt. On the otherhand, with the increase in 1) the number of scouts and 2) the ratio time-between-Home-melding/time-between-chats, much of the Map will not be in MotherShip (yet) but distributed among subpods. This argues for us providing mechanisms of local leadership, allowing subpods to combine their knowledge and assign subtasks.

When two scouts A and B chat they may each try to recruit the other to some Interesting discovery. Perhaps Sub A has found a vent chimney that is physically huge, is putting out large volumes of hot water, and is surrounded by indications of life. Sub A tells Sub B who might reply that he has seen an even more interesting eigenthing. They each may fail to recruit the other because the foundations of their evaluations may not be communicable.

Neural Nets constitute a sub-technology with attractive features. But the predictive wisdom captured, through experience, in a neural net seems not easily transferable.

Each individual of the pod will have its own face; that is, they will, by some mechanisms, be recognizable individually by other members of the pod. The "face" of a dolphinoid will probably include a voice, which could merely be an audible signature. Each sub's sonar can, by being individualized, serve as a signature.

Being able to recognize each other, scouts will share information about themselves. This will allow more efficient allocation of tasks: those scouts most capable with respect to some job, say the exploration of a particular kind of feature, are called first. Recognition of and storage of information about individuals produce, in each dolphinoid, his SocialUmwelt.

Chatting serves to keep up to date the SocialUmwelt of each pod member by the passing along of information about individuals. There is, naturally, a social component of the UberUmwelt.

Names of eigenthings allow efficiency in communication. Sub A can ask (in dolphinese) "Have you heard about the eruption at Vent 37" and B can avoid wasting time by saying "Yes." Efficiency in communication is more important as bandwidth is small - more important when speaking with sound than with the light of lasers.

There will be social roles. Roles differ in their SocialValue, the strength of the social mechanism to keep that role filled; or, more roughly: in each healthy animal society there are innate EigenRoles that are always filled. This, in some bees, is even true of the role of Queen: exactly one worker will assume queenly roles (some are pheromonal) after the queen dies. Our MotherShip has a recognizable voice. As the time-since-last-having-heard-MotherShip increases, a system of behaviors is motivated in TransponderSubs. These include commanding a scout towards MotherShip's last presumed location and calling on other subs for relevant information. Further, and of special interest here, each TransponderSub has an increasing drive D to become a PseudoMotherShip. If MotherShip continues to be missing, one of the subs experiencing the drive D, will be the first to have the role, a set behaviors, activated. One of those behaviors is of speaking ("singing"?) in the voice of MotherShip, thus returning the drive D to zero in all the other subs, insuring that there will be just one dolphinoid who changes roles. The PseudoMotherShip can, while energy reserves last, serve as a communications hub for more growth of the Map, and in assignation of duties. This example illustrates a method for dynamically allocating individuals to roles that does not require central control.

There will be divisions of labor within roles: this scout explores this area and that one that. There may be sub-roles within roles. Along with physical differences, ethorobotic training and Health will lead to individual differences in scouts, making some more appropriate for some functions than for others.

There may be different sizes of scouts, the smaller being more agile but of shorter range. These smaller microsubs will be better at maneuvering close to smaller objects of Interest and so be better for gathering data on animals. [I have had little to say about studying individual critters, thinking that for a very long time into the future, people will be needed for such discerning activity.] Some equipment may not be worth putting onto all scouts. Perhaps just one microsub in a pod will have a sub-bottom profiler (a kind of penetrating sonar) to make three dimensional models of vent mounds.

Structure of Chatting

This paragraph may be less interesting to you than to me, but these are my notes. We illustrate here how we might begin to design the social behavior of chatting.

The behavior of Chatting has the sub-behaviors InitChat and an EndChat. Chatting also has the components: Speaking, PausingAfterSpeaking, Listening, PausingAfterListening. These components allow for great flexibility in chatting; so that, if Sub A has a much higher drive than Sub B, A will talk in longer speech-segments and may speak two or more segments while B listens. We consider here only the case of exactly two subs chatting. When one is Speaking, the other is inhibited from Speaking. The duration of Speaking is drive-determined but finite and usually does not result in the speaking sub saying everything it has to say. Drive(Chat) is decreased by Speaking. Speaking by a Sub A is followed by PausingAfterSpeaking, the length of the pause being shorter as A's Drive(Chat) is large. While A is PausingAfterSpeaking, B is PausingAfterListening. The duration of B's pause is shorter as B's Drive(Chat) is larger. One of the subs will end a pause by Speaking before the other (this depends upon the relative values of the drives) and, so, inhibit the other from Speaking. And so on until one of the subs experiences its EndChat conditions.

Trailing Miscellany

There are two modes for the pod to be in, depending on whether energy will run out before time or time before energy. In an energy-conserving mode, bees send out a few scouts while the bulk of the hive rests. When swarming upon a threatening animal in a time-conserving mode, most of the hive rises to attack.

We, and many others elsewhere, have noted that some hints for designing social behavior of robots come from seeing that ant nests of impressive complexity and physical extent are built through the innate, small acts of individual ants. We have already begun following this hint: the

PodsModel is roughly analogous to an ant nest. We expect that a common language can be made to grow similarly.

We have had some fun here hinting at brave new social sciences. Our point of view has soared high. We now swoop onto some nitty-gritty tools for programming the first dolphinoid.

Calibrations and Play

Malfunctioning instruments can not only vitiate our scientific results, they can cause the loss of a microsub. The necessity to calibrate leads to an abundance of diverse behaviors for us to create. You can see here, if you try, another reason for the existence of "play" in animals.

The surface of a microsub can be altered by impacts and by adhesion of foreign material. A new optical reflection of the same strength as an old, may now be seen more dimly. A signal to the Rudder that originally resulted in traveling straight ahead may now cause a turn.

Our calibrations are of sensors and behavior. Of particular interest is behavior that constructs a ControlTable. If you have a particular interest in Fuzzy Logic, please do not skip TurnTable, Page I.17.

Sensors to be Calibrated

A calibration of a sensor S is a re-setting of S to make it give the same reading as a more accurate sensor or to match its own, earlier readings at some place where conditions are assumed to be constant. Calibration is done near and, in some cases, using HomeBase, possibly before every excursion from HomeBase and is surely done at Initialization.

Suppose S is a sensor that Sub can calibrate. There is the behavior Calibrate S and its specific motivation Drive(Calibrate S) which we here call D. D is a TimeDrive that increases towards its maximum of 1 on some schedule set by Mission Programmer. D is also increased by certain measurements that suggest S is reading incorrectly. High values of D can eventually lead to releasing, under appropriate conditions, the behavior Calibrate S (i.e. the appropriate computer subroutine) is executed. D may also influence other behaviors by feeding their drives (such as Drive(GoHome) ) and changing Confidence. If S is used in navigation, we might have planned that high values of D will prevent Sub from ranging deep into NavShadows.

Water temperature is read, while Sub is docked, and compared to the original at-dock reading and to the present reading of the HomeBase thermometer.

Distance and direction from Sub to HomeBase can both be measured using more than one instrument and more than one method. During Initialization, a physical marker is put in a permanent position P within HomeRange. Sub then and later, positions itself over this marker and measures the direction and distance to HomeBase. Sensing of range and of direction is also tested from positions other than P. When HomeBase sees the dolphinoid's laser and says, "You are 270 (west) of me, the dolphinoid had better see HomeBase at 90 (east) of itself. Other navigational sensors, in addition to the Compass, are tested.

HomeBeacon and HomeRanger are calibrated against instruments of shorter range but higher accuracy; instruments using higher frequency (hence shorter range) sounds or light.

Sub can pass over Home at a distance above it of HeightOne, the parameter used as the target height of Sub on the legs of the phase HSOne. Over Home, it tests its SonarSounder, a downward pinging sonar, and also its LaserSounder on a reflector on the top of HomeBase.

The water-clarity sensors are re-calibrated at HomeBase.

HomeBase makes special sounds so the "ears" of our dolphinoid can be calibrated. Amplitude and phase shift differences in two ears can indicate direction. Owls can locate a mouse by sound, striking the mouse even when the mouse is hidden under snow.

On what schedule is the calibration of a measuring system made? In ordinary, beginners' programming, one would decide on some fixed time or fixed count as a basis of an inflexible schedule. Ethorobotic's drive-based design allows the scheduling to flexibly depend upon numerous conflicting interests. The Urge to calibrate a slightly malfunctioning sensor of low priority can be much less than the Urge to calibrate a badly malfunctioning sensor of high priority.

Behaviors to be Calibrated

Sub will, from time-to-time, test its own physical performance. Two of the ControlTables that it maintains and uses to model and control its performance are the TurnTable and SpeedTable, used by RudderNode and ThrusterNode respectively. ControlTables will be partly trained before launch, then trained more fully during the Initialization Phase, and so on throughout the mission. To test (or to partly update) its TurnTable, Sub will go through many turns. Testing of the SpeedTable will be less conspicuous.

The various tables are stored in a non-volatile memory (e.g. a hard drive) and a version of them is stored in each computational node where used, if sufficient memory exists in that mode.

The less accurate a table is found to be in performance, the higher the drive to perform acts to "update" the table.

At Home, Sub makes straight-line runs at HomeBase, checking its speed indicators, calibrating its MaxSpeed and other values in SpeedTable. For accuracy in measuring range, direction and speed, laser devices are used both on the sub and on HomeBase, aimed at the sub.

Sub trims its roll and pitch for running straight ahead. (we cover, though not using the word, "yaw" in the next section).

Sub trims its density to equal, or slightly exceed that of local sea water.

By moving internal weights, Sub trims its attitude while SpeedInWater, its speed with respect to water, is 0. If facile enough, the trim-weights can be used to trim attitude while the sub is underway. Using the weights while running, rather than control surfaces (e.g. fins), can reduce energy-sucking drag, the resistance of the water to the passage of the sub.

Sub also re-measures at Home its maximum rate of climb and of descent, and, possibly of ascent and descent along a helical path with various settings of control variables.

Sub will, at HomeBase, measure the current from time to time, comparing the result with its previous measurements and those of HomeBase and other microsubs. Current will not be constant over time due to tidal effects. Even nearly constant currents passing cliffs and other obstructions can spin-off vortices. The Earth's crust is losing heat to the Arctic-cold water (at a particularly high rate in the regions of MissionOne), and this might contribute to a kind of slow "weather" involving vortices. Lava may flow infrequently, but when it does it surely stirs the water vigorously by heating it. Currents change the Heading required for moving in a desired direction. The amount of time spent on measuring current might depend on how variable, in absolute terms, is the current over time and over the terrain of exploration. But, this variability may not be known early in the exploration at some site.

The tables used for turning and accelerating contain fewer data entries than we will desire. Values intermediate to those entries can be computed using curve or surface fitting (also known as "interpolation") methods similar to those already used in computer graphics. A popular new tool for curve fitting is fuzzy interpolation.


TurnTable(Introducing Kinesthetic Learning)

The following begins a major foundation for intelligent behavior and shows how learning can refine an instinctive base. It also suggests application in this work of the currently popular tools of "fuzzy logic control". It may be too intensely technical for some readers.

The TurnTable, a combination of table and code, resides in RudderNode which has software for using it.

RudderSetting is a signal from RudderNode to its rudder actuator which (probably a "servo" motor) rotates and holds the rudder about its hinge to some angle. There is a RudderSetting of Neutral for traveling straight ahead. The Sub may be dynamically imbalanced so that "Neutral" may be a function of SpeedInWater, Sub's speed with respect to water. The value of Neutral may necessarily be learnable: the rudder can suffer wear and ding. A value of Neutral made incorrect by circumstances, or that is correct at some speed but not the present one, will result not in straight motion but curved. But it is true of every value of turning angle, not just 0, that its RudderSetting may be a function of speed and should be learnable.

During a turn there are "inertial" effects - both of physical rotation of Sub's mass and of moving the rudder from and then back to neutral.

When exactly in turning to a new Heading H, does the RudderNode command the rudder to return to neutral? If this happens when Heading has become H then Sub, by inertia will overshoot H. If the command is given too soon, then Sub will undershoot.

Another source of concern in controlling turning is that some devices (e.g. an ordinary compass) lose accuracy in turns.

For learnable accuracy in turns, our microsub will create and use a TurnTable.

Let partial TurnTuple mean a list of the numerical values of these variables:

(RudderSetting, SpeedInWater, TurnsDuration).

RudderSetting determines the angle through which the rudder is turned. Values of RudderSetting below Neutral will result in clockwise turning and, above, anticlockwise.

Let (r,s,t) stand for values of (RudderSetting, SpeedInWater, TurnsDuration).

To generate TurnTable, our dolphinoid goes through a sequence of partial TurnTuples (r,s,t) and, for each, performs a number of "trials". In each trial, Sub makes s its SpeedInWater and makes r its RudderSetting for the duration t. Sub measures the angle , the change in Heading, of the resulting turn. Using the average value of , the TurnTuple (,r,s,t) is recorded in the TurnTable and is there available for later use in navigation. A TurnTuple in the TurnTable is called a TablesTurnTuple.

Some part of the Brain, call one such part P for now, uses TurnTable to aid its own computations.

A hard turn of a short duration and a soft turn of a longer duration may produce the same change in Heading. So, our table does not represent, in itself, a one-to-one function. P must provide to TurnTable not only and s but one of r and t, say r, for specificity. If the TurnSignal (,s,r) is found lying within a TablesTurnTuple W, then TurnTable (the part that is code) returns W's value of t (TurnsDuration) to P.

A given TurnSignal (,s,r) [or a given (,s,t)] may not be found in any TablesTurnTuple. But, there will be triples in the table that are nearer this TurnSignal than are others. TurnTable "interpolates": it generates a TurnTuple lying "between" (in amongst) TablesTurnTuples that are near the given TurnSignal. Fuzzy interpolation is a currently popular method that can produce points intermediate to those in the table. Interpolation seems to be the main mathematical tool in real applications of "fuzzy logic control". If I am wrong in the perception, dear Reader, please correct me with evidence.

You may have noticed while docking a motor boat that the effectiveness of its rudder approaches zero as does the boat's speed. For maneuvering at low speeds we might have a bow thruster that can push left or right. The variable BowThruster results in some fraction of maximum possible power used (in the correct Sense, left or right) by the bow thruster. Perhaps BowThruster will have value 0 except when SpeedInWater is low, or there is an urgent need for a fast turn. Taking values of this BowThruster, our partial TurnTuples take the form:

(RudderSetting, SpeedInWater, BowThruster, Duration).

As simple as our model is so far, we are near to foundering on a computational reef. Let us estimate the number of entries in TurnTable. Suppose the number of values tested for each of the 4 terms of a partial TurnTuple are indicated by: (5,5,5,10). This may be a lower resolution than we would eventually desire, yet, this seems to give 1250 different partial TurnTuples. Each of these is to be tested in a number of trials and the results to be averaged to give an angle . Even if each TurnTuple thusly takes just a minute, the table would require more than 20 hours to be initialized. This argues for training the dolphinoid prior to the mission; though the resulting TurnTable need be neither complete nor as accurate as it will become with practice during the mission. We will plan for Sub to be skillful enough at birth to Dock and to avoid crashing into rocks.

During Initialization, Sub tests a selected set of existing TablesTurnTuples. Sufficiently large deviations in the predicted angle of turn will elicit re-training or even scrubbing of the mission. The mission can begin with an incomplete table; interpolation will proceed with whatever TurnTuples are available, though, perhaps, with some sloppiness. An outside observer might be perplexed to see a young submarine performing an occasional turn off a straight path while simply going from one point to another. We will know that our dolphinoid, rather than aimlessly playing, is training itself, improving its performance in turns. Scheduling of such playful experimentation is controlled via resolution of conflicts among drives. Training occurs near HomeBase because that is where information about position, velocity and so forth is most accurate.

Reviewing: our dolphinoid maps the Cartesian product of sets


to angle of turn. Selecting several sequences (r,s,b,t) the dolphinoid, during initialization, records in association with (r,s,b,t) the angular difference between Heading-before and Heading-after the turn. To use the resulting table, some node of the Brain may provide the desired angle of turn and a Duration. The table is used along with some constraints (not yet detailed) to compute values for RudderSetting and for BowThruster for the prospective turn. (We have noted that rather than Duration, a RudderSetting may be given; and, instead of either, some constraints may be given.)

The creation of the table implicitly takes into account the "inertia" of the rudder: the time it takes to reach a commanded setting and the time to return to its Neutral position.

Some node of the Brain will monitor the predictions of our tables, comparing them to actual results. Sub's turning "inertia" can change. Deviations of predictions from subsequent observations will be used by the HealthMonitor which can cause the release of corrective behavior (e.g. more practice in turning).

There are useful measurements we can add to TablesTurnTuples for no additional training costs. For accuracy in dead reckoning and in planning of paths, we might add the displacement (change of position) of the sub resulting from a turn. Under constant thruster power, a speed change will result from a turn: SpeedInWater will decrease.

MaxRollAngle of a turn can also be added to TurnTuples with no increase in training time. For some submarine designs, a turn can cause the nose to go down and the sub to descend. Our submarine will probably have a horizontal fin and a vertical fin (the rudder) which, seen from the sub's rear form the shape of a "plus". When the sub is in a turn caused by the vertical fin, it rolls about its fore-aft axis so that its top turns towards the interior of the circle of turn and its bottom towards the exterior. This rotates the vertical fin (the turned rudder), and rotates the thrust-vector of that fin off the horizontal, giving it a slight upward component; raising the tail and lowering the nose.

As you can see, the vertical and horizontal turning surfaces can interact. Accounting, via new terms of TurnTuples, for interactions among the surfaces that change the sub's direction of motion, can greatly increase training time. For the two surfaces in the case above (vertical and horizontal fins), an enlarged TurnTuple might have the form

(1, 2, RudderSetting, FinSetting, SpeedInWater, BowThrustHorizontal, BowThrustVertical, Duration1,Duration2). The training time will be much greater (half a year if we do not lower our ambitions or better use our heads). This increases the value of learning-on-the-job and of interpolation in an initially sparse TurnTable. A dolphinoid newly "born" may be ungainly like a puppy, then become graceful after much practice.

Training costs can to some degree be amortized over several microsubs. A second microsub, nearly identical to the first, can effectively, if just a bit awkwardly, begin life using a copy of TurnTable trained in the first.

In a planned later section "Predicting from Experience" we will compare "control" (of turns, for example) by "fuzzy logic", by "neural nets", and by other methods.

Other ControlTables

ThrusterSetting is a number that is a signal from ThrusterNode to the thruster motor. In training, our dolphinoid runs straight and level at various values of ThrusterSetting, noting the asymptotically resulting values of SpeedInWater, and also noting Power, the rate of energy flow into the thruster. To save time here, let us assume that a fixed duration ta is used, for all trials, to determine the "asymptotic" value of speed. The sequence (SpeedInWater, ThrusterSetting, Power) is stored in the SpeedTable. The ThrusterNode can subsequently know what ThrusterSetting to send to the thruster to give a desired speed and some other node can compute energy-efficient paths or time-efficient paths, depending on which is more important.

Example values of ThrusterSettings with fuzzy, friendly names: Slow = 30%, Cruise=70%, TopSpeed=100%,

SpeedsMinForTurning=20% (Below this speed, chosen arbitrarily here, the bow thruster may be needed for turning. These names show a connection to the currently popular ideas of "fuzzy logic control".

ThrusterNode in ordinary operation gets an initial value of ThrusterSetting for a DesiredSpeed from SpeedTable. After tatime has passed ThrusterNode compares SpeedInWater to DesiredSpeed. If the difference exceeds some threshold, ThrusterNode adjusts ThrusterSetting to increase or decrease the SpeedInWater toward DesiredSpeed. Actually, we can get much better performance by using the AccelerationTable (below).

The HealthMonitor records differences between power-setting as given by SpeedTable and the final power-setting that maintains desired speed. The HealthMonitor studies these deviations and calculates drives for corrective behaviors such as re-calibration of tables. More frequent and extreme (exceeding some threshold?) deviations lead more rapidly to reaction.

We could discuss the AccelerationTable, but I am sure that by now some Readers have gone to sleep or skipped on to another section. Avoiding detailed discussion of the AccelerationTable, we merely note, for your motivation, that the microsub will be able to accelerate rapidly to a speed by sending, for a brief period, more power to the thruster than will be needed to maintain that speed.

Another table can reduce oscillation when Sub is changing its Elevation.

We may ambitiously wish to provide for more interactions, such as exemplified by an accelerating, turning climb. Such complexities might, it seems, add years to training time. I do not know the extent to which this is avoidable, though I am sure we can be more clever about reducing training time than I have indicated so far. It may well be that a dolphinoid seasoned over years will be vastly more skillful than a yearling.

In creating the TurnTable, SpeedInWater is set equal to various values. To create the SpeedTable, the Sub must travel "straight-and-level". We have here what could be taken to be implicit references to three ControlTables each needing at least one of others before it can be created. This is not an insurmountable difficulty. Special control routines of lesser quality can be used for initializing the tables and then there can be a "boot-strapping" use of the partial tables to further fill them out.

Have you noticed that ControlTuples are used via an interpolation algorithm to relate an input vector X to an output Y, and that the parameters of a map are used via a filling-in algorithm to relate a 2D location X to an elevation Y (or some other quantity), and that, mathematically, these two operations are cousins? Heuristically, this indicates that they may both be subsumed under some more general mathematics. They differ from Neural Nets (which also produce a Y from a vector X) in that both ControlTuples and some map-parameters are points in the space of (X,Y) pairs but none of the "weights" in a Neural Net have such obvious meaning.


Future Topics

Future Topics

Here are some topics that I have thought about but whose notes are not presentable, or which I hold back for proprietary reasons. Perhaps this list will spark some thoughts and discussion. Some items are extensions of topics already introduced.

Other Missions: Clearing/Laying Mines

Chasing Turtles and Slow Whales and Giant Squid;

TomographySubs; Murky Deeps of the Deltas; DredgePrep;
Generalization Is What Neural Nets and Fuzzies Do
Principles of Behavior:
Instinct and Learning in die Umwelt der Dolphinoids
Shaping Behavior (Trial and Reward)
Predicting Performance from Experience
Software for Nodes; Internodal Variables.
Navigation: Transponders (long, short, very short baselines) vs Beacons
Path-Shaping by the Method of Dynamic Waypoints
Confidence: A Vector Impacting Drives
Close Study (of seafloor eigenforms)
Feature Following: long eigenforms (connecting the dots)
Images from Echoes and Vision from Lasers
Sonar Imaging at WHOI
Laser Imaging at HBOI and at ART
Pattern Recognition via Eigenforms
Sea Station: a Movable Buoy-Link to NASA
Problems and Difficulties, When to Abort
SpiderSubs lay optical fibers.
Using Pyrobenthic Sounds
Building a Dolphinoid or a MiniDoid; RoboBoats
Dolphinoids of MIT, WHOI, HBOI/FAU
Fuzzy Logic Control: Fire behind the Smoke?
Multiple Sensors Reduce False Alarms: as a product of two numbers near 0.
Drives and Behaviors, more examples and structure.
Eigenforms and WorldModel (Eigenwelt)
Generating ControlTables


Go to Top of this Document     Top of this Site

1) Smart Robotics with Benthic Applications

2) Knowledge Preservation via Netbooks


BOTTOM of Document