Stochastic Actions
PROBLEM: Action effects are stochastic
- Actions transform state-distributions to state-distributions
- Maximize “probability” of goal satisfaction
- Plan assessment itself is hard
APPROACHES:
- Conjunctive planners
- BURIDAN [Hanks et. al., 95] uses UCPOP techniques to put candidate plans together and assesses them
- Multiple /redundant supports
- Disjunctive planners
- Pgraphplan [Blum & Langford, 98] modifies Graphplan to support some forms of stochastic planning
- Forward search; value propagation; Envelope extension
- Maxplan [Majercik & Littman, 98] uses EMAJSAT encodings to solve stochastic planning problems
- Chance variables & Choice variables. Equivalence classes of models that have the same values of choice variables. Find the equivalence class with maximum probability mass.