Incremental Recompilation for Standard ML of New Jersey
CARNEGIE-MELLON UNIV PITTSBURGH PA DEPT OF COMPUTER SCIENCE
Pagination or Media Count:
Probabilistic planning methods can have various objectives usually they either maximize the probability of goal achievement or minimize the expected execution cost of the plan, and therefore assume that the agent that executes the plan has a risk-neutral attitude. Although there are many situations where risk-sensitive behavior Is more appropriate, researchers have largely ignored the question how to incorporate risk-sensitive attitudes into their planning mechanisms. Utility theory shows that it is rational to maximize expected utility, given that the agent accepts a few simple axioms and has unlimited planning resources available. Thus, researchers might believe that one could allow for risk-sensitive attitudes by replacing all costs with their respective utilities for an appropriate utility function. We show that this is usually not the case and, moreover, that - in general - the best action in a state can depend on the, total cost that the agent has already accumulated that is, the Markov property does not need to hold. However, we demonstrate how one can transform planning problems for risk-sensitive agents into equivalent ones for risk-neutral agents provided that exponential utility functions are used. The transformed planning problems can then be solved with probabilistic AI planning methods or, alternatively, dynamic programming methods.