Cox, M. T. (2007). Perpetual Self-Aware Cognitive Agents. AI Magazine, 28 (1), 32-45.

@Article{Cox2007,
  author = {Cox, Michael T.},
  title = 	{Perpetuaal Self-Aware Cognitive Agents},
  journal = {AI Magazine},
  year = 	{2007},
  volume = 	{28},
  number = 	{1},
  pages = 	{32--45}
}

Author of the summary: Janine Fitzpatrick, 2009, jfitzpat@connect.carleton.ca

Cite this paper for:

• Agents need to learn from past performance failures by determining which factors led to the error and avoiding them in the future. [p32]
• Case-based reasoning consists of using past experiences for problem solving and/or explanation purposes. [p36-37]
• To understand the world "is to interpret the world with respect to the knowledge and experience one (human or machine) currently has in memory." [p43]
• Acute awareness of the world involves being able to understand when the world needs to change and to form a goal to accomplish this. [p43]
• Self-interpretation involves generating learning goals in order to improve "knowledge used to make decisions in the world" and to be able to "explain ourselves to others." [p43] Requires understanding of the self.
• SYSTEM: INTRO

Few AI systems are able to use comprehension of their environments to derive goals.

Components of INTRO include Meta-AQUA, which can reason about itself via introspective multistrategy learning, and PRODIGY, the planning component that makes the action sequences for INTRO. [p33]

Meta-AQUA: Analyzes its own reasoning failures to improve its performance of story-understanding. In order to do this, a multistrategy approach is used. A comprehension method is chosen (such as case-based reasoning or script processing), and explanations for 'interesting' input are added to the model of the story. [p33]

Abstract explanation patterns (XPs) describe the initial explanations for interesting events. If an alternate explanation is offered in the story, an 'explanation failure' is triggered and "passes a trace of the prior reasoning to the learning subsystem." [p34]
Three processes must occur for effective learning:

1. Explain the failure (developing a meta-explanation pattern [Meta-XP] by probing the memory for the failure symptom and retrieving an XP)
2. Decide what to learn (Meta-XPs either generate learning goals or additional questions)
3. Construct a learning strategy for set of learning goals [p34]

Meta-AQUA has comprehension proficiency because it can represent both the story and its own cognitive processes as it relates to the story. The latter ability is necessary to reason about factors behind explanation failures. The system can use the learning goals generated by Meta-XPs to change its background knowledge through a process called case-based introspection. [p35]

The AQUA system within Meta-AQUA is the performance component which applies XPs to explain stories. Meta-AQUA explains explanation failure and is thus able to understand itself. [p36]

The generative planner of this architecture, called Prodigy4.0, achieves the given goal state through a four-step decision process:

1. Goal is selected from list
2. Operator to achieve goal is selected from list
3. Object bindings are selected from list
4. If applicable, a forward chain (apply operator) or backward chain (continue subgoaling) is applied [p36]

Prodigy/Analogy uses the Prodigy4.0 generative planner as well as a case-based planning mechanism. [p36]
It uses derivational analogy, meaning that it reconstructs lines of reasoning between old and new cases. By creating justification structures, Prodigy/Analogy can "provide a principled mechanism that supports an agent's awareness of what it is doing and why." [p37]

While Prodigy/Analogy uses its successes to learn, Meta-AQUA uses its explanation failures.

INTRO: can declare goals which "provide intention and a focus for activities." [p38]
Uses memory of earlier percepts and actions, as well as interpretation program, to map an output action choice from an input. [p38]

There are four components to the INTRO agent:

1. Perceptual subsystem - means of input; translate representations of Prodigy/Agent and Meta-AQUA. [p38-39]
2. Effector subsystem
3. Understanding subsystem (Meta-AQUA) - explanations are used by INTRO to create a goal that can change the environment (problem recognition). Meta-AQUA uses both 'Goal Monitor' (with, for example, knowledge goals) and 'Memory Monitor' (represents memory retrieval and storage) windows. [p39]
4. Planning subsystem (Prodigy/Analogy) - actions generated to change the physical environment; "in response to an active environment, INTRO generates its own goals to change the world." [p41]

What distinguishes between requirements of learning goals (changing knowledge to avoid repeating reasoning error) and achievement goals (achieve an alternative state of a 'flawed' world)? The current system can generate learning goals when it understands that its knowledge is flawed; achievement goals are automatically generated and a decision process to distinguish between the two goal types remains to be implemented. [p42]

"...A metacognitive integration between Meta-AQUA and PRODIGY remains unfinished." [p43]

Summary author's notes:

• Page numbers above refer to the pages from the printed version, not the published version.