Anthrobotics™

Table of Contents

Background

About The Anthrobotics Approach

Summaries of Portions of the Technology

    Creating "Awareness" In An Autonomous Decision System

        •  Prediction
        •  Relevancy
        •  What-If Trial Actions
        •  Look-For System
       •  Efficient Computation

   Natural Language

        •  Translation
        •  Determining Meaning
        •  Thought Elements

    Primitives" In Data Structures

       • Prediction Factors
        • Recognition
        • Simulation/Prediction

    Relevancy Nodes

    Same Parallel System For Prediction As For Relevancy

    "Feelings" Generation

    "Metabolism" Feelings

    Face-Driving System

    Story-Telling System

    Language Production

   "Meta"-Interface System

    Sociality

    Learning

        • Cognitive Learning
       • Relevancy Learning
       • Motor Learning
        • Learning From Language


Background

Building intelligent entities, especially autonomous entities and languaging
entities, is obviously a difficult thing to do. Many brilliant and
well-funded teams have worked at this problem for a long time. And such
teams have developed many tools which will be useful in the eventual
building of such an intelligent entity. Some few examples of such tools are
parsers, expert systems, neural nets, vision systems, fuzzy logic, robotics
and control systems, and face recognition systems.

There has been a need for a new overall approach to using some of these
hard-won tools in a unified way in the building of an intelligent entity,
especially more-general-purpose such entities. The thrust of the
Anthrobotics technology is to provide this kind of new overall approach.
Anthrobotics provides new viewpoints and tools in problem areas such as
representation, categorization, relevancy, feelings, awareness, learning,
and natural languaging.

About The Anthrobotics Approach

The Anthrobotics approach and proprietary technology has many new and
interesting ideas. One of the many core ideas is to provide representations
(e.g., for "objects" and for "situations") along a concrete-abstract scale
in such manner and of such type that a straightforward inclusion computation
may provide answers as to whether any particular "concrete" is or is not
included in any particular "abstract". This permits an entity, for example,
after sensing/determining external information about the more-concrete
circumstances it is in, to compute whether and how its such circumstances
have relevance to it as measured by their inclusion within one or more of
the more-abstract situations which do have relevancy to the entity (e.g.,
nodes on a plan).

Another core idea is to provide basic categorizations within the
representation system which have (at least at higher levels) essentially a
one-to-one vocabulary tie-in with natural languages.

A third core idea is that whatever system of categorization and/or
representation an intelligent entity uses, it must be useful at least in a
predicting of the immediate "concrete" consequences of the immediate
"concrete" circumstances as known by such entity.

Favorable feasibility testing of Anthrobotics solutions on a demonstration
software platform has been done in the above core areas and many others.

                           Go to Table of Contents

Summaries of Portions of the Technology

Creating "Awareness" In An Autonomous Decision System

Whether an "aware" ADS is operating in a "real-world" setting or in a
computer or virtual world, it must have sensors (or sensor-like abilities)
with which to apprise itself at least of nearby potential events. And it
must be normally able to make a mapping representation (also called
"structured situation") of its specific circumstances/environment.

Prediction

Further, an autonomous decision system (ADS), even if presented with an
adequate mapping representation of its specific circumstances/environment,
its "structured situation", should also understand the implications of its
circumstances; it should be trying to, or able to, predict at least the
immediate futures, the consequences, of its present situation.

                           Go to Table of Contents

Relevancy

And before it may be said to be "aware", the ADS should also know, or be
trying to know, its organism-environment ("O/E") "relevancies" within its
present situation and within its predicted future situations. Thus, the ADS
should have a relevancy system, preferably its hierarchically-organized
representations of its problems and plans, and its plans including its
subgoals and goals. And its present and predicted circumstances (within its
O/E transactional "field") should be constantly compared to its relevancy
system for the ADS to understand the relevancies in its present
circumstances. It would seem that, if anything may be called "awareness" in
an autonomous decision system, it would be the CONTINUOUS and SIMULTANEOUS
operation of prediction and relevancy functions, of implication and meaning
functions, providing, in its transacting with its environment, an
often-chaotic interactive prediction/relevancy "field" for making choices.

What-If Trial Actions

Furthermore, the ADS should have a what-if system within which it can
internally "try out" its proposed decisions, its trial actions, to predict
the relevancies, the alternative consequences, of its proposed decisions and
choose which decision to take in its real world.

Look-For System

And the ADS should have a "look-for" system whereby information about what
things to best then look for with its senses may be transmitted to
sensorimotor areas from its prediction/relevancy field. Continuous
assessment of relevancies, of threats and opportunities, of positions and
status of plans, etc., permits informing the perception system of "what
things to look for". Since, by their nature, most relevancy systems are made
up of abstract situations defined in part be abstract objects, the
abstract-object representations of the active part of the current plan, for
example, may be saved; and then a dictionary look-up with some easy
computation tells what percepts to "scan for". For example only, the ADS may
contain a table of all of its known concrete-object representations, which
are defined as sets of the Anthrobotics primitives (elsewhere herein further
described). Searching may then provide a list of all of these
concrete-object representations which contain the primitive(s) of the
abstract objects being looked for. From such list may be taken all instances
of perceivable primitives (or recognizable characteristics) compiled by
number of occurrences; and then a minimum of one (more if desired)
highest-frequency such characteristic become(s) the percept(s) "scanned
for".

Efficient Computation

Furthermore, the system of internal representation of the ADS should permit
efficient computation in all of the above areas. Such efficient computation
requires that the categorization/representation system selected permit easy
computation of "whether a concrete situation/object is included in an
abstract situation/object".

                           Go to Table of Contents

Natural Language

Further, it would seem that a useful way to look broadly at natural language
is to see it functioning as "substitute experience" in its ability to
present a mapping representation, a structured situation, in interaction
with a human or ADS. From this viewpoint, "understanding the meaning" of
natural language is metaphored by having an ACTUAL experience and assessing
the meaning of this actual experience from what happens in the human's
body/brain systems (or ADS systems) responsive to this experience, i.e.,
what happens with prediction/relevancy "awareness", with emotion, with
associations and memories, with metabolism, etc. Thus, for our described
ADS, natural language meanings/understandings depend upon language evoking
responses/occurrences in (e.g., "mammalian-type") non-linguistic substrate
systems of an autonomous decision system. Thus, in a digital system, the
data-structure types and the data within each type are what must be
primarily communicated from one ADS to another in order to do such evocation
and have "understanding". Ideally, two digital ADS's may transfer such
information instantly in parallel; but humans do not have that ability and
must make such transfers serially (and slowly).

Of course, in modern human communication, these data are often mixed and
overlaid with approval-disapproval language, logic language, etc., all of
which must be dealt with for full communication abilities. Still, for our
ADS, natural language may be roughly seen and processed as (1) a series of
serial clues, e.g., language "snippets", to inform as to what data-type is
coming in what location in the statement/sentence, and (2) natural-language
words/phrases corresponding to data of the type "expected" from such clue.
And since the language receiver (among humans) has about the same kinds of
data-types and data as the sender, such communication often works. ADS's, by
using our data-types and our clues/snippets and a human natural-language to
correspond to specific data and for forming such clues (and synonym lists
for both the snippets and the data), may communicate in natural language
with humans.

Translation

In fact, translation may be seen as the use of a first natural language for
giving data to an ADS and the use by the ADS of a second natural language in
reporting its internal data to a human native speaker in the second natural
language.

Determining Meaning

The best way to determine natural language "meaning" is to provide an
appropriate autonomous decision system and an interface system which
transforms and uses elements of (a first) natural language to produce
modifications and stimulation in the non-linguistic elements and systems of
the substrate ADS. Then, when desired, the ADS can produce (if desired, a
second) natural language showing the "meaning" which has been "transmitted"
by the natural language input to such interface system. Natural language
"syntax" problems arise when there can be no direct simultaneous
transmission of all data-type/data available from one ADS to another.

Thought Elements

So non-linguistic elements of the ADS (the data in the data-types
mentioned), non-linguistic representations, preferably bear certain
relationships to natural language, relationships of the type permitting,
e.g., learning by analogy, metaphor, example, etc. And so such
non-linguistic representations (which are hereinafter sometimes called
"thought elements") should preferably bear precise similarity relationships
among them, as for use by learning programs and natural-language use.

Thus, the selection of thought elements should ideally provide: efficient
"awareness" (as set out briefly above and hereinafter) computation both in
prediction and in relevancy areas; efficient "fit" with human knowledge and
learning in areas like "similarity" and "metaphor"; efficient use
cross-culturally and in translation and other compatibilities with human
thought and languaging; efficient data storage of minimum elements;
efficient "computability" among data elements; and efficient hardware and
software "parallelism", etc. The most important of these thought elements
are below discussed.

                           Go to Table of Contents

"Primitives" In Data Structures

It is preferred to start with representing a set of categorization
"primitives" with which to begin the task of categorizing the infinite
variety of the world and usefully construct a finite representation of it,
hopefully resulting in a relatively small set of such primitives and in a
method of categorization comparable to that of humans and at the least
comfortable to and understandable by humans. What is desired to start with
is a way to categorize "things" or "objects" (in that this is the chief way
that humans begin to break down the world).

Prediction Factors

For our ADS, one primitive to be selected should be a "prediction factor"
(PF) in that its form of expression should be such as to be efficiently
computable, e.g., by simulation methods, in predicting the consequences of
the object (having the PF) within a mappable representation of objects, a
structured situation (Sx). As a preferred digital data structure, an Sx may
be seen as a set of event rows (Ex), each event row including at least: (1)
a "concrete" object representation (Px), (2) a location representation (Lx),
(3) a movement or vector representation (Vx), and (4) at least one "doing"
or bodily-movement representation (Bx or BFM). [For information about how to
arrange further descriptions in this or similar areas, please contact Marty
Stoneman (marty@indirect.com)]

Recognition

Also, importantly, "recognition" of a "thing" or "object" by an ADS having a
sensory system depends (again, briefly and roughly) on (1) a 1st
categorizing (by Recognition Factors, RF's, each such factor specifying a
selected range within the output of each piece of sensory equipment focused
on the "thing" or "object", for example) of a "concrete" object or "thing"
based on the sensory information coming from a small space (the location of
the "thing"), representable as part of a geometrical situation, a structured
situation, during a small time period; (2) a 2nd categorizing [by Prediction
Factors, PF's] of a "concrete" object or "thing" based on the tendency(s) of
the object/"thing" to behave, through time, in certain specific ways in
certain kinds of specified relational situations, depending upon the 2nd
categories (types) of other "local" things in said geometrical situation
[which other local "things" act/react to the 2nd-categorizing type of said
"thing"]; and (3) given a set of said 1st categorizings of a "thing", the
assignment, on the basis of said set of said 1st categorizings (using, for
example, a look-up table), of a set of said 2nd categorizings (by PF's) of
said "thing". Until PF's ("hierarchied" or ordered in a useful order) have
been assigned to each event representation of a presented structured
self-situation, there has not been complete "recognition". The set of RF's
and the set of PF's so assigned to a thing will sometimes be hereinafter
called its "concrete" object representation.

Thus, for a thing/object to be called "recognized", one must be able: to
incrementally predict about the thing in the relevant geometrical situation
which includes said thing, in the framework of the total set of 2nd
categorizings of things; and to determine the "meaning(s)" involving the
thing (as it is, within its specified local geometrical situation), in the
framework of the total set of 2nd categorizings of things. The act of
"recognition" may be defined, according to the present invention, as the act
of assigning a "sufficient" representation to the object/thing. A
"sufficient" representation (which makes an object "concrete") is one which
has sufficient information to enable a participation by the object
representation in a simulation (as herein exampled) with its current
environment or mapping representation for predicting the immediate
consequences of the object-and-situation.

                           Go to Table of Contents

Simulation/Prediction

A simulation system of our type may be used not only within an autonomous
decision system, but also as the "engine" to "run" a reality simulation or a
"virtual" reality or as an engine for a game, etc.

Relevancy Nodes

For an ADS to be "aware", it must do more than run predictions; and one of
the other things it must do is to compare such predictions (and its
recognized situations that are the start of the predictions, of course) with
a relevancy system to begin to determine its threats and opportunities. For
our ADS, a relevancy system includes a hierarchied set of problems and
hierarchied sets of plan-strings associated with each problem. Each problem
"node" and each "node" of each plan include a representation of an
inclusional situation, an Si (also sometimes called herein a "self-relation"
or a "relational situation"), in the same form as the inclusional situation
of a PF (i.e., in the form of a boolean set of SE's, as more specifically
described below). Each such "node" identifies at least one "abstract"
situation or Si whose specific relevancy depends upon the position of its
node in the hierarchically-organized set (and upon other factors associated
with a node, as hereinafter described).

Same Parallel System For Prediction As For Relevancy

And the described form of the representation permits the same parallel
processing system to participate in the "fast-forward-simulation" of the
current representation to compute its immediate consequences (using
"default" ADS self-actions or self-inactions). I.e., the process of
simulation/prediction may use the same parallel processing system that is
used to determine a current representation's relevancy-inclusion in the
relevancy-planning set of nodes. In each case, the main computation is to
determine any inclusion of a structured situation in a relational situation,
i.e., whether an Sx is included in an SE (and, eventually, in an Si, which
may be composed of a boolean set of SE's).

In doing parallel processing, if each parallel processing module represents
a specific genetic or learned "sub-evaluation" (SE), object representation
to be tested for inclusion in the SE) may be tested in parallel in each such
processing module and any inclusion(s) or "hit(s)" stored. Then each
specific space-time geometry between/among specific objects from the
viewpoint of each object [so that all relations r are included] may be
tested in parallel in each such processing module. Thus each processing
module could have the parallel, simultaneous ability to test a given SE from
each of many viewpoints [the number of viewpoints will usually, at least for
most perceptual systems, be a low number, probably 5 to no more than 20].

Each processing module which "satisfies" both requirements, i.e., the
general relation has "hit" between "hit" objects/ Px's, sends a signal to
each appropriate "summary" processing module (one for each Px viewpoint),
which keeps updated on all combinations of SE's which are "interesting"
Si's; and each processing module reports on Si "hits" from its viewpoint to
its decision processing module (which could be the same processing module as
its summary processing module). Each decision processing module, which keeps
updated on all decisions for any Si's (including any "hierarchical" rules),
selects a decision for its viewpoint Px in this situation. One (or more)
decision processing module(s) is a problem/plan processing module for
keeping track of decision-making and what-iffing, etc. When a new Sx is
recognized, or prediction decisions/actions of all Px's of the Ex's are
reflected in a new SxT, the described process is started again,
appropriately.

To make all such SE processing modules "the same", they would have to
include general relation-testing submodules as well as general set-inclusion
submodules (as well as a "viewpoint" and "hit" memory). And they would then
be temporarily or permanently adapted to test for a specific SE (e.g.,
"firmware"). The capability for some on-the-spot modifications would be
helpful for implementing learning, etc. As mentioned elsewhere, "fuzzy
logic" or "neural nets" or other methods could be involved in the work of
the summary decision-choice steps, if desired, after the "hit" SE's have
been determined.

                           Go to Table of Contents

"Feelings" Generation

Our ADS may include a feelings system for generating and using
representations of various simulated feelings or "emotions". A set of
registers is provided for the below purposes, and the "amount" of a
particular feeling will be roughly proportional to the number in such
register. For most of the feelings registers of the ADS, the incremental
additions (in numbers) to such registers will be in terms called "arousal"
and the incremental subtractions will be called "relief".

Although the source of the various arousal incremental additions is varied,
the source of the relief incremental subtractions is a quantity proportional
to the "muscle action" or movement of the ADS (with a relatively fixed
"basal-metabolism" quantity). For example, if the emotion means of the ADS
provides, for use by the autonomous decision system, incremental
representations of "fear" in amounts essentially hierarchically ordered
according to inclusions of the mapping representation Sx in the hierarchical
set of n problem representations, there may be, say, an increment of 100 (on
a scale of 0-255) of arousal generated as an incremental amount of "fear"
and going to the fear register (until replaced by the next fear increment);
and this increment is added to a "total arousal" (or "excitement") register
(which register also receives arousal increments from other sources, as will
be described). If, at the beginning of an increment time, the total arousal
register is at 200 (on a scale of 0-255), and if for that increment fear of
100 and other arousals of 80 are added to the total arousal register, and if
for that increment, a relief decrement of 200 is subtracted from the total
arousal register, then, for the next increment, the total arousal register
starts out with a reading of 180.

Further, for example, the ADS may include emotion means for providing, for
use by the ADS, incremental representations of "hopelessness" depending
essentially upon whether, in the operation of the planning means, in the
plan-set for the highest hierarchical problem representation in which the
mapping representation (Sx) is included, such mapping representation is
included in none of the subgoal representations associated with said
plan-set; and a hopelessness register (similar to the fear register) is
provided for incremental storage.

Further, for example, the ADS may include emotion means for providing, for
use by the ADS, incremental representations of "disappointment" depending
essentially upon whether, in the operation of the predictive planning means,
when it is predicted that a "future" such mapping representation will be
included in a next subgoal representation, such "future" mapping
representation is in fact not included in a next subgoal representation; and
a disappointment register (similar to the fear register) is provided for
incremental storage.

Further, for example, the ADS may include emotion means for providing, for
use by the ADS, incremental representations of "surprise" depending
essentially upon whether, in the operation of the predictive planning means,
when it is predicted that a "future" such mapping representation will not be
included in a next subgoal representation but rather one subsequent thereto
(i.e., closer to the "goal" of eliminating the current problem), such
"future" mapping representation is in fact included in a next or subsequent
subgoal representation; and a surprise register (similar to the fear
register) is provided for incremental storage. However, the contents of this
register will be deducted rather than added to the total arousal register so
as to increase pleasure (reduce pain) as described shortly hereinafter.

During any one increment of time, the number in the total arousal register
will be an additive function of the incremental arousals of all such
specific arousal registers like those described. And, in this case, the
total arousal register will be an indicator of the general feelings of
excitement/calmness of the ADS depending upon the highness/lowness of the
register contents. Also, other arousal registers may be used for other
purposes, for example, an unpredictabiity arousal increment when either (1)
an object can not be "recognized" or (2) a predicted situation does not in
reality occur (as in surprise and disappointment scenarios, etc.). This
unpredictability register reading is useful in signaling when cognitive
learning should be initiated, as mentioned below.

The rate of change of the number in the total arousal register from time
increment to time increment may be used as a rough indicator of
pain/pleasure of the ADS depending upon the highness/lowness of a
rate-of-change-of-total-arousal register. That is, the faster total arousal
is going up, the more pain; and the faster total arousal is going down, the
more pleasure.

                           Go to Table of Contents

"Metabolism" Feelings

Computations from simulated (or real, from robot sensors) metabolism, may be
made from appropriate incremental additions to appropriate registers
representing such things as amount of hurt or hunger or tired of the ADS;
and from appropriate incremental subtractions from such registers
respectively of healing-done or food-eaten or rest-had.

Face-Driving System

Also, once the amounts of various feeling-representations are determinable,
these may be used to "drive" a human-face-representation, e.g., by using
drawing formulas for the various lines of the face (e.g., eyes, eyebrows,
eyelids, mouth, forehead creases, nose, cheek creases, etc.) and providing
formulas linking the amounts in the various feelings registers to the
various face-draw formulas. In this manner, the "true" emotions/feelings of
the ADS may be observed by others, similarly to human interaction.

Story-Telling System

Further, the ADS may be given a system for telling "true" and "interesting"
stories about the experiences of the ADS within its virtual or real worlds.
Human-type storytelling (and "episodic memory") may be efficiently simulated
by saving data according to the following techniques: (1) at an appropriate
"high" (selected) level of pain, to start the story, save the time and place
and the "concretized" current problem situation/node (i.e., for the
structured situation which is included in the relational situation of such
problem node, save the included concrete objects of the structured situation
along with the node information); (2) save as the next story element the
current strategy and the first plan node found to be included, along with
the concretized relevant data and the action being tried to overcome the
problem (and any "high" feelings at this time and at any further node
report); (3) then save such information about the next plan node reached and
any further action to be tried and any what-iffing predictions; (4) then
continue to so save about any next plan node reached and any further action
to be tried and any what-iffing predictions; and (5) end the story about
reaching a goal when there is a "high" (selected) level of pleasure. The
story should preferably report (following the "rules" of the selected
natural language) in the concrete terms of the structured situations, not in
the abstract terms of the relational situations.

                          Go to Table of Contents

Language Production

For this story-telling use of the ADS, or to have the ADS answer questions,
or for other language production uses, the language production program or
module must of course have access to the pertinent thought elements (data)
representing the material to be reported. Then, for, and depending upon, the
specific natural language in which the report is to be made, the ADS must
have access to (1) the normal order of reporting, both within sentences and
among sentences; (2) the "snippets" or phrase structures which signal to a
listener that the included word(s) represents a certain kind of data; (3)
usual connectives; and (4) the word(s) which represent the data to be
reported. For variety, the ADS should have access to synonymous or alternate
acceptable ways to use such orders, snippets, data-word(s), etc.

It is noted that other technologies, e.g., parsing or logic or speech
production or prosody or speech recognition, may be combined with the
instant invention where desired.

"Meta"-Interface System

For ease of use in designing objects for "worlds" for simulation computation
according to this invention, or for fast entry of appropriate "defaults" in
using incoming language to evoke the ADS, it will be most efficient to be
able to assign numerous PF's to an object (and to "lock out" from assignment
other PF's). For this purpose, according to this invention, a
"meta"-interface system may be used in which the user selection screens have
natural-language categories of behavior, etc., which may be selected for
each object; and such selection at the top user level enforces the desired
specific PF assignments/lock-outs to be made. Such meta-PF "buttons" may be
"cascaded" in multiple levels to efficiently permit selection of even many
dozens of PF's with one top-level selection.

Sociality

The ways in which the ADS' feelings and social interaction are systematized
may provide "empathic" maximum "user-friendliness" between humans and the
ADS's of the present invention. Briefly, to begin to achieve such sociality,
the ADS should be given, as a problem very high in its hierarchy, the "goal"
of avoiding situations where it is not "copying" (like imitation, empathy,
agreement, etc.) with the humans in its immediate company. The learning
which will occur in this problem area will be of the "cultural" type so that
the ADS will "fit in" to the "culture" or "civilization" in which it exists
and will be under pain to learn the rules of such culture.

Friendship and Love

Also, the ADS should constantly compare its behaviors under specified
circumstances with those of each human (or other humanoid ADS) around it and
fix for each such "other" a number (say, from 0 to 100) for the relative
number of behaviors in common with it; the ADS will thus determine which
others are the most the "same kind" as the ADS and in what hierarchical
amounts and order (called herein the "K-Number" or "K#" assigned to each
such "other" by such comparison program). Then the K# will be used so that,
in the ADS interrelating with a specific other, the amount of fear
associated with such aforementioned "not-copying" problem will vary roughly
according to such K#. Thus, to the ADS, a K# of, say, 50 might represent a
strong "love" relationship, and a K# of, say, 25 might represent a
friendship; while a K# of, say, 8 might represent an alien or stranger.

Learning

Learning by an ADS may be seen as the addition of new thought-element data
or the modification of existing thought-element data. For learning, there
must be: (1) an internal signal that the current situations (the "now" as
herein defined) require learning; (2) the saving or ability to save the
"now" data for use in learning; (3) the kinds of defined-similarity provided
by the thought elements of this invention, so that learning can proceed in
an orderly manner beginning with small changes in previous most-similar
data; (4) ways to indicate that learning of an appropriate nature has
occurred; and (5) a learning procedure or program appropriate to the kind of
learning being attempted.

An ADS should "save" and store temporarily, for possible use in learning, a
selected number of Sx's (structured situations), say about the previous
(from current) eight seconds worth. Thus, for example, if about three Sx's
are perceived each second, there will be (for a selected saving-period of
eight seconds) 24 Sx's always temporarily so saved.

And the fast-forward simulation program used by the instant predictive
planning system may be used to compute and so temporarily save, say, about
the next eight seconds predicted to occur. Thus, for example, if about 10
Sx's are predicted-computed each second, there will be (for a selected
saving-period of eight seconds) 80 consecutive "future" structured
situations always so saved. When the "16-second window" comprising the "now"
may be herein mentioned, it refers to this continuous temporary saving of
"now" data for storing and use whenever a mentioned learning signal occurs.

                           Go to Table of Contents

Cognitive Learning

Cognitive learning in an ADS has to do with (1) maintaining appropriate
assignment of PF's (and RF's) to thing-types or Px's for use in
"recognition", and (2) modifying existing PF's or adding to the store of
PF's. The signal that cognitive learning should be done (using the current
or "now" data to be saved upon such signal) will be either (1) inability to
perform "recognition" of something during a construction of an Sx, which
signal may be used to add to the feelings system as an "unpredictability
arousal" contributing by addition to total arousal and thus pain/pleasure,
or (2) inadequate prediction (see the flowchart of Fig. 10) by the system
predicting/simulating the "future", discovered by regularly comparing
predicted Sx's (for a selected time-frame) with actual perceived Sx's in
that time-frame, which signal may also be used to add (to the feelings
system) unpredictability arousal. It is noted that, in the predictive
planning system, failure to real-time achieve a predicted plan-node status
is a similar signal as for unpredictability arousal and also may form the
basis of a separate "disappointment arousal". Similarly in the case of
surprise.

The cognitive learning procedure for an ADS constitutes experiment, either
by external "playing"/experimenting in the real world or by internal
experiment in which possible new PF data is "tried" in the prediction system
to measure whether (and how much) the data substitution helps to make the
predicted outcome equal the saved actual outcome. And the order of such data
changes to be attempted will be from most similar to gradually less similar
data, reminding that in the PF structures of the present invention, the
degree of similarity is defined (by the relative "distance" from
"sameness").

Relevancy Learning

Relevancy learning in an ADS has to do with (1) addition of problem nodes or
(2) addition of plan nodes. The signal that relevancy learning should be
done (using the current or "now" data to be saved upon such signal), will be
either (1) an unexpected (unpredicted) pain or fear in an amount exceeding a
preselected threshold, signalling an unexpected problem situation (and thus
requiring a new problem node), or (2) an unexpected (unpredicted) pleasure
or fear-lowering in an amount exceeding a preselected threshold, signalling
an unexpected problem alleviation (and thus requiring a new plan node). The
relevancy learning procedure for an ADS constitutes using the saved "now"
experience [in the form of a short series of concrete situations or Sx's
just preceding an unexpected pain or pleasure (or fear or fear-lowering)]
and data-similarity considerations to write an appropriate new node, to be
tested by its adequacy in further experience.

Motor Learning

Motor learning in an ADS has to do with (1) selection of a set of serial
motor actions to be learned and (2) practice of such ordered set of actions
until the initiation of the set produces a smooth, coordinated finishing of
the set. The signals/occasions for motor learning (using the current or
"now" data to be saved upon such signal), will be either (1) the attempt to
copy the actions of a nearby other of the same "kind" as the ADS (i.e., the
ADS has a relatively high K# for the other), signalled by "not-copying
arousal" in proportion to such K#, or (2) the attempt to externally
experiment (or "play") with a thing in a non-coordinated way required by
cognitive learning, using the signals and procedures of such cognitive
learning.

Learning From Language

According to our technology, in which natural language is seen as evoking
substitute experience, similar ADS learning as set forth above may occur as
a result of natural language. For example, language such as analogy and
metaphor provides the "now"-data for the ADS and points (usually) to the
similar cognitive data to be changed by experiment. Also, for example,
language such as images and stories (involving "empathy" or "identification"
of the ADS with a certain character) provide the ADS with experiences of
pain and pleasure at certain points of "now"-data, permitting relevancy
nodes to be changed by such experience. Also, for example, the natural
language items or stories which the ADS "wants" to remember (for more
"copying" with others, for instance) may be practiced as motor actions
(called by us in this case "membering") until they may be "re-membered" by
initiation of such motor action "string".

                          Go to Table of Contents

                            © 1996 Anthrobotics™
 

  Copyright© 1996-2006 Anthrobotics

This site is owned and operated by Anthrobotics. No material from  Anthrobotics or any Web site owned, operated, licensed or controlled by Anthrobotics may be copied, reproduced, republished, uploaded, posted, transmitted, or distributed in any way, without the prior written permission of Anthrobotics, except that you may download one copy of the materials on any single computer for your personal, non-commercial home use only,  provided you keep intact all copyright and other proprietary notices.

Modification of the materials or use of the materials for any other purpose is a violation of Anthrobotics copyright and other proprietary rights.  Attributed extracting for comment or criticism is permitted.