PRODIGY
A diagram of this architecture
Authors
Jaime G. Carbonell, Craig A. Knoblock, and Steven Minton
- Uniform representation language --> PDL: Prodigy Description Language(FOPC-like)
- Multiple learning methods
- Inductive & analytical learning
- Learns from failure as well as success.
cf)SOAR learns from only success
- Glass box hypothesis --> Each PRODIGY module is visible to the other modules and share data with them.
cf) SOAR is taking 'black-box' philosophy where the internal structure of chunks is not open to inspection or modification
- Hierarchical abstraction planner(problem solver)
- Search control : depth-first means-end analysis in default
- Top-down automatically generated monotonic(no alteration of higher level) abstraction (ALPINE)
- Difficult part of the plan is considered in higher abstraction level
- Dependency-directed backtracking <-- By combining learning and problem solving
- All the modules have:
- Common representation based on first order logic
- Global knowledge sources(domain knowledge, control knowledge, problem solving trace, etc.)
- Utility problem --> EBL can degrade performance if costs and benefits of learned knowledge ignored.
Utility = (AvrSavings * ApplicFreq) - AvrMatchCost
- Factors that contribute to performance degradation
- Low application frequency (i.e. overly specific)
- High match cost
- Low benefit
- Deliberative learning --> PRODIGY decides what to learn about, when to learn it, and why it learned it.
---> Directs its learning to maximize utility, and perform selective forgetting
cf)SOAR: reflexive learning
- Architecture-level axioms --> Theory that descibes the problem solver
- Domain-level axioms --> Theory that descibes the task domain
- Deliberative reasoning
- Maximum rationality --> Intelligent agents will behave so as to maximize the likelihood of achieving their goals
- A model of idealized rational intelligence unlike SOAR which models human cognition