czwartek, 8 sierpnia 2013

Reading Gary Drescher "Made Up Minds"

The "schema mechanism" system developed in Gary Drescher's PhD project reminds me of Anticipation-Based Learning Classifier Systems, but it is more AGI / cognitive-architecture worthy because of its representation building facilities. The book is very novel for its year 1991, before modern RL theory became popular in AI. ABLCSes are quite recent development in LCSes (although with an early publication in 1990). I have (good) sentiment towards Learning Classifier Systems, they were my first encounter with a more cognitive form of AI, very long ago.

It would be cool to redo Gary Drescher's project, but using Rich Sutton and Hamid Maei's recent results -- Gradient Temporal-Difference Algorithms with off-policy learning -- instead of schemas.

The principle behind Gary's project is constructivism, the opposite of nativism. Almost all structure of the world, even what Kant claimed to be necessarily needed a priori -- binding of experience into objects having features -- can be learned from input by a relatively simple mechanism. You might think that therefore constructivism isn't only radically against Chomsky, but also David Deutsch'es ideas I've been quoting recently --  bashing of empiricism and of logico-positivism. But consider this: the construction algorithm, to succeed, must have universal reach in David Deutsch'es terms. And Gary Drescher accepts the criticisms against logico-empiricism, "Even the most rudimentary conceptions of the physical object cannot be defined by the schema mechanism as any function of the sensory primitives." Section 8.6 stresses "Why Non-naive Induction Must Be Built In"; and the mechanism needs to solve similar problems with counterfactuals. The system uses counterfactuals for learning and concept invention. "The difficulty and obscurity of the concept of counterfactuals is, I suspect, a reason that its fundamental importance for learning systems has been late to be recognized, rather than a reason to consider it an implausible basis for learning."

Notes below focus on chapters 3 and 4 as these describe the mechanisms of the cognitive system. I pick a couple of nuggets from later chapters. Chapter 2 presents Piagetian theory, elaborating on initial stages of development. It is worth reading in full.
  1. Conditions are conjunctions of items. Primitive items are sensory facts and synthetic items are beliefs.
  2. Primitive items are binary: On / Off, synthetic items are ternary: On / Off / Unknown.
  3. Objects and relations between objects are supposed to be stable configurations of schemas, synthetic items and composite actions.
  4. Schemas are identified by a triple: preconditions, action, postconditions.
  5. Primitive actions change state of the world and agent.
  6. Composite actions are identified by a postcondition they achieve.
  7. Therefore a schema might be a refinement of the postcondition of a composite action under given preconditions.
  8. Accessible value in a state is the maximum value achievable along a reliable path from the state.
  9. Instrumental value is assigned to items along reliable paths to a goal. I.e. items in a state have its accessible value as instrumental value. Instrumental value is transient.
  10. Delegated value is assigned proportionally to (1) the difference of average accessible values of an item: when the item is On minus when the item is Off; and (2) the duration of the item being On. It is more permanent.
  11. Item of frequently instrumental value only has delegated value when it is not "readily accessible" -- readily accessible items have the difference in accessible values (between their On/Off states) close to zero.
  12. To avoid runaway of propagating delegated value through cycles, the value propagated is half of the delegated value.
  13. Attention via hysteresis-habituation loop: recently selected schemas have more weight but decreasing with consecutive selections.
  14. Reweighting to upsample schemas with rare actions.
  15. Promote actions with inverse effects (turns an item On/Off right after it was turned Off/On) as it heuristically leads to reliable schemas.
  16. Marginal attribution:
    1. Start from bare schema: {}-A->{}; add results in spinoff schemas for items whose positive-transition (from Off to On) or negative-transition rates following action A are higher than their averages for all actions.
    2. Add (negated) items in contexts of spinoff schemas which (anti)correlate with validity of the schema, making the spinoff schema more reliable.
    3. All statistics count only unexplained transitions and correlations -- when the items aren't in results or contexts of existing valid schemas. (Statistics are reset after a spinoff accordingly.) Statistics are only collected by most specific schemas accounting for a situation. This increases sensitivity to regularities and reduces combinatorial explosion.
  17. Schema chaining requires that all items of the (primary) context of a following schema are provided by results of preceding schema. It's used for composite actions, etc.
  18. Schemas have associated extended context (and extended results). Extended contexts and results are mutable (evolve over time). Besides being a data structure for spinoff formation, extended context adds to the condition for activation of a schema. Schema chaining cannot rely on the corresponding spinoff schemas, because it often requires general primary contexts.
  19. For a composite action, besides conditions that need to hold initially, there can also be conditions that need to hold throughout the action.
  20. Synthetic items are designed to identify invariants when all apparent manifestations change or cease (compare Piagetian conservation phenomena).
  21. Keep track of local consistency: the probability that a schema will be valid right after it has been valid, and expected duration of consistency: how long since onset of validity to first invalid activation. (Recall attention via hysteresis.)
  22. A synthetic item is a reifier of validity conditions of an unreliable locally consistent schema, called its host schema -- action called probing action and result called manifestation. I.e. it is the state item such that when added to the context (precondition) of the schema, if the schema were activated, the action would bring the result.
    1. The intention is that the synthetic item captures a persistent feature, like presence of an object, while the remaining context items of the host schema capture transient features, like effector configuration.
  23. Learned verification conditions set the state of a synthetic item:
    1. Host schema trial: when host schema is activated, On resp. Of if it succeded resp. failed.
    2. Local consistency: the state remains as changed for at most a period of expected duration of host schema's local consistency (for On, local "inconsistency" for Off). Then revert to Unknown.
    3. Augmented context conditions: the extended context of the host schema (which collects evidence from spinoff schemas).
    4. Predictions: "If a synthetic item appears in the result of a reliable schema, and that schema is activated, then in the absence of any evidence to the contrary, the mechanism presumes that that schema succeeded".
  24. Above mechanism approximates a synthetic item, but the synthetic item is not coextensive with any function of cumulative inputs (i.e. of the inputs history).
    1. "The schema mechanism grounds its synthetic items in the reification of counter-factual assertions; the subsequent adaptation of its verification conditions is driven by that grounding."
  25. Composite action is created for each spinoff schema that has a novel result.
  26. "A composite action is considered to have been implicitly taken whenever its goal state becomes satisfied [...] Marginal attribution can thereby detect results caused by the goal state, even if the goal state obtains due to external events." Which together with hysteresis leads to imitation.
  27. Backward Broadcast mechanism and action controller learn proximity of schemas (results) to goal states (of composite actions). If reliable chains of schemas are found, they are incorporated into composite actions. The chains are also used for forward prediction.
  28. Action controller handles special cases: indeterministic actions (schemas with same contexts and actions but various results), repetition, on-the-fly repair (detecting schema that make applicable some component of an interrupted action).
  29. Schema with composite action cannot spin off a schema with part of the composite action goal in the results.
  30. There is no "problem resolution" mechanism. Rather, some schemas hit a dead-end, and are taken over by schemas that capture more fruitful regularities.
  31. I haven't understood how inversely indexed representations (synthetic items) work. (par. 6.4.4)
  32. Note that synthetic items do not represent identity of for example tactile-visual objects. This isn't bad though because the system's mistakes reproduce Piagetian errors at corresponding developmental stages. Errors mean unreliable schemas which leads to further development.
  33. Now on to more far-fetched stuff. "The new conception [learned abstract concept] reifies the set of circumstances under which a piece of one's computational machinery behaves a certain way."
  34. "Consciousness requires knowledge (and hence representation) of one's own mental experiences as such; the schema mechanism does not come close to demonstrating such knowledge."
  35. Unimplemented mechanism: subactivation. "To subactivate an applicable schema is essentially to simulate taking its action, by forcing its result items into a simulated-On state (or, if negated, a simulated-Off state)." The simulated states are entirely distinct from actual states and all mechanisms are duplicated for them. But statistics are shared, and spinoff schemas are created "for real".
    1. Simulations are serial but parallel chaining search will cache the knowledge.
  36. Unimplemented mechanism: explicitly represent inverse actions to make them available for subactivation (i.e. simulation).
  37. Override generalizations: when a derived (by simulation) schema prediction is wrong because a direct schema from which it was derived is overriden, derived schema should be overriden too, without penalty for the derivation. A new schema will be created to capture this exception.
    1. "The suggestion is that deductive-override machinery may permit the schema mechanism to escape the fallacy of naive induction. The key is to regard the conflict between a reasonable generalization and an absurd but always-confirmed generalization as just another conflict between generalizations expressed at different levels of description."
  38. "The reason [not to build in] a variable-matching implementation of generalizations, is just that there is no apparent way to support such an implementation without abandoning the constructivist working hypothesis by including domain-specific build-in structure. [...] Perhaps the system itself could be designed to devise explicit structured representations to support variablized generalizations. [I]f virtual generalization fails, devising such machinery may be vital to the schema mechanism."
  39. "A schema's extended context is essentially a connectionist network solving a classifier problem."
  40. Unimplemented: clustering (i.e. hierarchical modeling); "having coordinated coarse- and fine-grained spaces mitigates the combinatorics of showing the path from one fine-grained place to another, because the path can be represented as a coarse segment to get in the right vicinity, followed by a fine-tuning segment."
  41. Unimplemented: garbage collection. Schemas: not contributing to goal achievement, not spawning new spinoffs, seldom activated, perhaps even those that are activated but recreation opportunities are more frequent than activation opportunities.

Brak komentarzy:

Prześlij komentarz