An Empirical Comparison of Incremental Search Algorithms for On-Line Planning

Introduction In an on-line planning problem an agent must select and execute a sequence of actions to maximise some objective criterion, where the time and resources used in action selection count in assessing overall solution quality. Such problems are representative of a broad class of bounded rational (Russell & Wefald 1989) real world planning problems, such as control of autonomous vehicles and control of chemical plants, where an agent must perform a task as well as possible within its limited physical and computational abilities. Thus, to maximise overall performance the agent must tradeoff the cost of resources used in planning against the possible benefits of improved action selections. Incremental search (also called agent-centred search (Koenig 2001)) is one way of performing this trade-off, where actions are selected based upon an heuristically guided partial search through the space of possible future action sequences. As partial search provides incomplete information about the best action, action selection becomes a problem of decision-making under uncertainty. Meta-control can thus be achieved by trading-off the benefits of more complete information for action selection against the computational cost of performing additional search. This paper presents a comparison of three incremental search algorithms which use different meta-control strategies; fixed (Korf 1990), myopic Expected Value of Computation (EVC) based (Russell & Wefald 1989) and hyperopic EVC based (Baum & Smith 1997). The aim being to clarify the algorithms’ strengths and weakness. This can be used to guide the development of more flexible alternatives; for example by using multiple levels of abstraction.