symbolic_cognitive_robotics.md

Symbolic cognitive robotics - Design notes

Author: Jean-François Cloutier

Research project: Symbolic Cognitive Robotics, Active Inference Institute

Last updated: April 10 2024

Design notes on the next iteration of a cognitive architecture for Lego robots.

Formative concepts

• Active Inference (an agent actively minimizes surprise to survive)
• Enactivism (an embodied agent's perceptions and actions are constructively co-dependent)
• Apperception (predictive sense-making comes from the discovery of unified causal theories)
• Mortal semiosis (sense-making and sign-acting are intrinsically grounded in the agent's drive to survive)
• Society of Mind (an agent is animated by a collective of cognition actors interacting with each other and the world)
• Constraint Closure (the cognition actors constrain how the Society of Mind as a whole can change, and vice-versa)
• Kantian Whole (the parts -cognition actors- exist for and by means of the whole -the Society of Mind-)

Society of Mind (SOM)

• An agent grows and evolves a Society of Mind from its engagement and experiences
• A SOM is a "connectome" of Cognition Agents (CAs)
• Each CA has an umwelt consisting of other CAs or some sensed aspects of the agent's environment
• A SOM starts with a priori constituents
  • The a priori objects are self, world and ground
  • The a priori relations is in
    • objects ground and self are in world
  • The a priori value domains are
    • boolean - [true, false]
    • integer ranges, including percentage
    • colors - [unknown, black, blue, green, yellow, red, white, brown]
  • Detectors and effectors are exposed as a priori CAs
    • Each defines an a priori vocabulary of
      • beliefs (e.g. color, distance, touched)
        • with value domains (e.g. colors, 0..100, etc.)
      • actions (e.g. spin, reverse spin)
  • Metacongnition actors have a priori capabilities and each CA has a priori, shared introspective belief voacabulary (see below)

Exteroception vs interoception vs introspection

• Exteroception => Cognition of (more or less abstracted) sensations from the external world
  • e.g. the distance to an obstacle
• Interoception => Cognition of (more or less abstracted) bodily sensations
  • e.g. a motor is stalled
• Introspection => Cognition of sensations about computations by Cognition Actors (CAs)
  • perceived by Metacognition actors
  • e.g. a CA has low predictive capability
• A CA can non-exclusively coopt other CAs into its umwelt as sources of interoceptive or exteroceptive sensations, and as effectors of actions

The umwelt of a Cognition Actor (CA)

• The (immediate) umwelt of a CA is a (small) fixed set (fixed at instantiation) of other CAs
  • The CAs in an umwelt also have CAs in their own umwelts
    • All the way down to a priori CAs
  • The transitive umwelt of a CA is the CAs in its immediate umwelt plus the CAs in their transitive umwelts
  • The level of abstraction of a CA is the maximal depth of its transitive umwelt
  • A CA builds its umwelt from CAs of lower and uniform levels of abstraction
• CA Interfaces
  • A CA's interface is what it exposes to other CAs, namely
    • The vocabulary of its beliefs (what others can make predictions about), composed of
      • extant (observed), latent (imagined) and synthetic (derived) objects,
      • extant, latent and synthetic relations/properties
        • a latent or synthetic property is always boolean-valued
    • The actions it affords
      • either a priri actions (e.g. spin and reverse spin of motors) or synthetic actions
  • All CAs have a common vocabulary of meta-cognition beliefs
    • exposed to meta-cognition actors
• CA events
  • A CA emits events listened to by the CAs tp which unwelts it belongs, and by the meta-cognition actor overseeing it
    • when prompted by predictions about its beliefs, a CA emits
      • prediction errors given its currently held beliefs
        • with varying precision
        • from perceiving other CAs
        • or from cognitive self-assessments/beliefs (from introspection)

A CA's perceptions

• A CA processes perceptions one discrete time slice after another

  • the time slice duration is constant for the CA and proportional to the CA's abstraction level
    • i.e. the depth of its transitive umwelt

• Perceiving is making predictions about about the beliefs of CAs in its umwelt

  • possibly corrected by prediction errors emitted by CAs in its umwelt

• Perceptions are

  • Uncontradicted predictions
  • Prediction errors can be emitted in response to predictions, with attached precision
    • If multiple CAs respond to a prediction with prediction errors
      • The prediction error with the highest precision has sway
      • Tie-breaking is random

• The precision of a prediction error (0% to 100%) is a function of:

  • The confidence of the emitting CA in the contradicting belief, which is a function of:
    • The accuracy of the supporting causal model behind the belief (see below)
    • The duration of the perceptual trend supporting the belief, modulated by (see below)
      • the average precision and variance of the perceptions aggregated by the trend (see below)

A CA's synthetic beliefs

• A belief is what's imagined (latent) or synthesized by the CA
  • from its perceptions
    • which are unrefuted predictions + prediction errors about the beliefs of CAs in its immediate umwelt
• Beliefs are available to other CA's as synthetic or latent and thus novel perceptions
• Beliefs are abduced predicates
  • needed to formulate a causal theory (latent)
  • or needed to label significant perceptual trends (synthetic)
• Beliefs have associated normativity (pleasant vs unpleasant vs indifferent beliefs) from ambient feelings
• "Thin now" vs. "thick now" beliefs
  • Thin now beliefs are
    • unobserved but imagined/abduced properties/relations/objects to (causally) make sense of observations
  • Thick now beliefs are
    • synthetic beliefs, induced from, and thus supported by, perceptual trends

Perceptual trends support synthetic beliefs

• Perceptual trends (from the analysis of past perceptions) support synthetic beliefs
• A trend is about the relation or property of a given object
  • A relation associates two objects
  • A property associates an object and a value from a defined domain of values
• A trend is itself given a value
  • a trend is either stable, unstable, up, down
    • trend(<predicate name>(<object name>), <trend value>, <since>)
    • for relations, up/down describes a count of related objects,
    • for properties, up/down describes the rise/fall in values
      • property value domains are ordered from lesser to greater
• Memorizing trends
  • A CA keeps a limited history of past perceptions
    • It is the CA's short-term memory (developing trends)
    • Individual, past perceptions are eventually forgotten
  • A "compressed" trend (a trend about now forgotten perceptions) can be preserved as long-term memory
    • compressed(<trend>, <time interval>)
    • It is associated with remembered past beliefs (that the trends supported)
      • When past beliefs are forgotten so are their associated compressed trends
    • Compressed trends represent the CA's long-term memory

Inducing beliefs from trends

• Beliefs are induced from the analysis of trends, themselves produced from analyzing a perceptual history.
  • A belief induced form the analysis of a trend is then supported by that trend.
  • Induction can result from
    • Associating
      • Synthetic properties/relations are supported by attention-worthy (strongly felt or surprising) trends
        • <synthetic property name>(<object_name>, true | false)
        • <synthetic relation name>(<object name>, <object name>)
    • Partitioning
      • Parts-whole beliefs are induced by detecting boundaries in an observed object.
        • in(<new object name>, <object name>)
      • How are boundaries detected?
        • An object has differentiable, stable sub-trends that coincide in time
        • This might indicate that different parts of the object were being observed at different times
          • e.g. "patch of food" in the "ground" in the "world" ("self" is always in the "world")
      • A part is not of the same object type as the whole (assuming no fractal objects)
    • Categorizing
      • Beliefs about partition cause the abduction of new objects (the parts)
      • The "part" object is assigned a new (abduced) object type
      • is_a(<object_name>, <new object_type>)
• A trend is significant and worthy of belief induction if
  • it breaks surprisingly from a previous trend
  • or if correlates with a change in feelings

Trends, feelings and the normativity of beliefs

• Normativity (something being good, bad or indifferent) is always about trends
  • It exists in the "thick now"
• Normative valuation comes from associating trends with feelings (see below)
• A trend takes its (normative) value from the intensity of concurrent, ambient feelings
  • trend_value(<trend>, good | bad | neutral)
• A belief supported by a trend takes the normative value of that trend
  • A belief associated with a bad feeling is unpleasant, else it's pleasant (good) or indifferent (neutral)
  • Since trends have lengths, the normative values of trends have duration
    • e.g. a long-lasting unpleasant belief are worse than a short-lasting one

CA actions

• Changes in properties/relations observed by a CA are either caused by latent processes or by actions.

  • In a static environment, they are caused entirely by actions!
    • No perception without action and no action without perception

• To make sense of/apperceive the consequences of actions, they must be observed together with the property/relation changes they (may) cause

• A CA exposes, by name, the actions it can execute

• A CA must always be capable of acting

  • i.e. it has at least one effector CA in its transitive umwelt

• The action repertoire of a CA consists of

  • the actions it synthesized
    • starting from the distinct actions exposed by CAs in its umwelt

• The CA of an effector exposes atomic actions

  • For example, a wheel CA exposes the atomic actions "spin" and "reverse spin"

• A CA syntesizes actions from the actions exposed by CAs in its umwelt, names them and exposes the former in turn

• A synthetic action is a named list of actions

  • e.g. action_2 = [action(effector_ca1, action_1), action(effector_ca1, action_2), action(effector_ca2, action_3)]
    • an action can be repeated to synthesize a replacement, synthetic action
  • a synthetic action is, via closure, a sequence of atomic actions

Why does a CA defines a new action?

• A CA synthesize a new action because a sequence of actions is empirically associated with a significant belief change
• Belief changes that may lead to action synthesis are
  • Abduced object, property or relation from a causal model (thin now belief)
  • Correlation with a belief-supporting trend starting/ending/enduring (long now belief)
    • The sequence of actions that runs before/through the trend is extracted
  • Babbling
    • A CA synthesizes an action to see what would happen if executed
      • As a variation on an action already in its repertoire
        • Amplify sub-sequences via action duplication
        • Tone down sub-sequences by reducing duplication
        • Splice and recombine a synthetic action

Action intents

• An intent names an action that a CA wants executed.
  • A CA can intend any action in its repertoire
• What motivates action intents by a CA (from less to most motivated)
  • Babbling
    • to maybe cause a "chance" belief
  • Evidencing
    • to impact confidence in a belief (thus the precsion of reported prediction errors)
  • Eliminating
    • to terminate an unpleasant belief
• A CA intends at most one action per time slice
  • It intends the most motivated action in its repertoire
    • favoring, but not always, actions of the most successful policies (see below)
  • If multiple actions are considered
    • A motivation tie is randomly broken

Action execution

• Execution of an intended action is inhibited if another CA concurrently intends an action that
  • covers it (is a super-sequence)
  • or is identical and has higher normative motivation
• All actions taken are observable by all CAs
  • The atomic actions from the closure of synthetic actions are observed
    • During time slice T of the CA
    • If a sub-sequence of the observed atomic actions recreates a synthetic action in the repertoire of the CA
      • then the longest synthetic action is what is observed, plus the second longest etc.

CA action policies

• A policy is an action associated by a CA with a belief, a goal (verification, elimination) and a success rating from its executions

Feelings

• Feelings are agent-wide signals about the satisfcation of self-maintenance requirements
• Feeling types
  • Fullness
    • How full/empty are energy/resource stores (hunger when low)
  • Integrity
    • Degree of structural integrity (pain when low)
  • Engagement
    • High when learning (boredom when low)
• feeling(<feeling type>, good | bad | neutral)
• Motivational ranking
  • Fullness > Integrity > Confidence > Engagement
  • The agent dies when energy/resources are depleted
  • The agent is immobilized when integrity is too low
• Feelings are centrally computed from
  • detector sensations
    • touch - integrity decreases
    • color - resources increase if color == food type
  • effector sensations
    • work done - energy decreases
  • CA cognitive sensations
    • mental effort - energy decreases
    • prediction success rate - competence increases/decreases
    • relevance (rate of received predictions, intended composited actions)
  • The passing of time
    • healing - integrity increases
    • base metabolism - resources/energy decreases
• Any change in fullness/integrity intensity is signaled to all CAs
• For each CA, for each time slice, there's an average intensity of each feeling type

Constraints

• Umwelts (when closed) must be acyclic directed graphs but not necessarily trees
• A CA must not include a CA in its immediate umwelt if the latter is already in its transitive umwelt.
• A CA must be either cognitive or meta-cognitive, never both
• Only one synthetic action in a conflicting set can be executing at any given point in time
  • A synthetic action conflicts with another if their closed sequences have any simple action type * incommon.
  • Practically speaking, only one synthetic action is allowed to execute at any time
• A CA must not remove an element from its interface if it is used by another CA
  • to formulate a causal theory
  • to synthesize a belief or action
• A CA must archive a belief (without normativity) or action and its (compressed - abstracted ) support when the support is gone but the belief or action is still used by other CAs
• A new belief must not be created if its support is subsumed by the compressed support of an archivedbelief
  • The archive belief is ressucitated and given the current support
• An archived belief/action must be deleted if the belief/action is no longer used by another CA.

Initial state of the Society of Mind (SOM)

• Initial CAs
  • One primitive CA per effector (wheel_1, wheel_2)
  • One primitive CA per detector (color_sensor, touch_sensor, obstacle_sensor, beam_sensor, etc.)
  • One meta-cognition CA with as umwelt all the primitive CAs *Initial steady state variables (sources of feelings)
  • Integrity 100%
  • Energy 100%
  • Foresight 100%

Meta-cognition CAs (MCAs)

• Every CA of level N belongs to the umwelt of one MCA associated to that level
  • The level of a CA is the number of edges from the CA to primitive CAs
  • For a Society of Mind (SOM) with N levels, there must be one MCA per level 1..N, plus one MCA for level N+1 with an empty umwelt
  • Once a CA is added at level N + 1, an MCA is immediately created for the empty level N + 2 etc.
• An MCA observes only cognitive sensations from its umwelt

Meta-cognitive actions

• An MCA is exploring the "connectome space" of CAs at one level of abstraction, looking for a beneficial organization
• An MCA at level N can
  • Create a CA at level N
    • And add level N-1 CAs to its umwelt at creation
  • Remove a CA

Cognitive sensations

• The cognitive sensations are
  • effort as
    • apperception engine usage
    • memory load
  • foresight as
    • prediction success rate
  • stability as
    • rate of change in beliefs, actions, causal theories
  • relevance as
    • rate of received predictions,
    • percent of actions composited by other CAs