S. Russell and E. Wefald (1991). Do the right thing : studies in limited rationality (Chapter 2: Metareasoning Architectures), MIT Press

@Book{Russell91,
  author =       "Stuart Russell and Eric Wefald",
  title =        "Do the right thing : studies in limited rationality",
  publisher =    "MIT Press",
  year =         "1991",
  series =       "Artificial intelligence",
  address =      "Cambridge, Mass",
}

Author of the summary: David Furcy, 1999, dfurcy@cc.gatech.edu

Cite this paper for:

• SYSTEMS: TEIRESIAS, MRS and SOAR (a brief discussion of each as a metalevel architecture).
• Intelligence = acquiring and using declarative knowledge and gradually compiling it into an efficient execution architecture.
• Metalevel knowledge can be derived from domain-level knowledge and domain-independent knowledge about the internal operations of the system.
• Taxonomy of knowledge types.
• Taxonomy of execution architectures.
• Compilation as a necessary process for increased efficiency.
• Importance of the coexistence of all types of execution architectures at both the object-level and metareasoning level with compilation processes to transform one type of knowledge into another.

Summary

Detailed outline

Metareasoning

• Traditional definition: Reasoning about reasoning i.e., reasoning about entities that are internal to the system, as opposed to object-level (or OL) reasoning, which manipulates symbols referring to external entities.
• In this book: There exists an identifiable decision procedure to choose computation steps (Note that declarative representations are not required). Metareasoning occurs when the nature and order of computation steps in the OL decision procedure depends on current events (not hardwired at design time).

• Notion of introspective fidelity, when the recommendations of the metalevel are always followed by the OL reasoning. Most current systems (except SOAR) do this.
• Heterogeneous versus uniform metalevel architectures. The latter enable infinite regress. This problem explains why optimal control of ALL reasoning is not possible and justifies the need for default, hardwired and/or approximate reasoning.

Execution Architectures

• Static knowledge:
  • description of some aspect of a world state -- condition(state)
  • action results -- condition(result(action,state))
  • utility of world states -- utility(state)
• Dynamic knowledge:
  • Type A: condition(s) -> condition'(s)
  • Type B: condition(s) -> condition'(result(a,s))
  • Type C: condition(s) -> utility(s,v)
  • Type D: condition(s) -> best(a,s)
  • Type E: condition(s) -> utility(result(a,s),v)
  • Type F: condition(result(a,s)) -> best(a,s)

• Type D or condition-action rules can be compiled using type B and F knowledge (among other ways).
• Type E or action-utility rules can be compiled using type B and C knowledge.
• Type F or goals can be compiled from type C knowledge and the decision-theoretic principle.

• Decision-theoretic systems: A + B + C + DT
• Goal-based systems: A + B + F
• Action utility systems: E + DT
• Production systems: D

Existing metalevel architectures

• TEIRESIAS: Rules infer the absolute or relative value of OL rule applications but the metalevel lacks a model of the OL decision procedure.
• MRS: The metalevel selects both tasks and the methods to accomplish them and contains a first attempt at modeling the probability and utility of the outcomes of computational steps at the OL.
• SOAR: The metalevel consists of conducting full-scale simulations of OL computational steps and of chunking the results into productions. The simulations causes SOAR to be inadequate for a bounded rational agent.

Summary author's notes:

• none