Eric B. Baum 1 NEC Research Institute, 4 Independence Way, Princeton NJ 08540 eric@research.NJ.NEC.COM Abstract The point of game tree search is to insulate oneself from errors in the evaluation function. The standard approach is to grow a full width tree as deep as time allows, and then value the tree as if the leaf evaluations were exact. This has been effective in many games because of the computational efficiency of the alpha-beta algorithm. A Bayesian would suggest instead to train a model of one's uncertainty. This model adds extra information in addition to the standard evaluation function. Within such a formal model, there is an optimal tree growth procedure and an optimal method of valueing the tree. We describe how to optimally value the tree, and how to approximate on line the optimal tree to search.
The efficient use of limited computational resources is an essential ingredient of intelligence. Selecting computations optimally according to rational metareasoning would achieve this, but this is computationally intractable. Inspired by psychology and neuroscience, we propose the first concrete and domain-general learning algorithm for approximating the optimal selection of computations: Bayesian metalevel policy search (BMPS). We derive this general, sample-efficient search algorithm for a computation-selecting metalevel policy based on the insight that the value of information lies between the myopic value of information and the value of perfect information. We evaluate BMPS on three increasingly difficult metareasoning problems: when to terminate computation, how to allocate computation between competing options, and planning. Across all three domains, BMPS achieved near-optimal performance and compared favorably to previously proposed metareasoning heuristics. Finally, we demonstrate the practical utility of BMPS in an emergency management scenario, even accounting for the overhead of metareasoning.
Sequential decision problems are often approximately solvable by simulating possible future action sequences. Metalevel decision procedures have been developed for selecting which action sequences to simulate, based on estimating the expected improvement in decision quality that would result from any particular simulation; an example is the recent work on using bandit algorithms to control Monte Carlo tree search in the game of Go. In this paper we develop a theoretical basis for metalevel decisions in the statistical framework of Bayesian selection problems, arguing (as others have done) that this is more appropriate than the bandit framework. We derive a number of basic results applicable to Monte Carlo selection problems, including the first finite sampling bounds for optimal policies in certain cases; we also provide a simple counterexample to the intuitive conjecture that an optimal policy will necessarily reach a decision in all cases. We then derive heuristic approximations in both Bayesian and distribution-free settings and demonstrate their superiority to bandit-based heuristics in one-shot decision problems and in Go.
Each one of these problems can be approached and solved using optimizing or satisficing techniques. Each stage involves complex tradeoffs that can be addressed off-line or online. For example, using a more precise model of the environment may complicate the problem definition and may force the system to compute less precise answers to the problem. The key question is whether bounded optimality is a useful approach to all or some of these problems. In other words, the question is whether there are any advantages to making optimal decisions within an approximate model, rather than making approximate decisions within a more precise (or even perfect) model.