Commentary on Baum’s “How a Bayesian Approaches Games Like Chess”

Contents

The relevance of the paper to the symposium lies in its use of decision-theoretic metareasoning to select nodes to expand and to terminate search. A computation is relevant if it promises to provide utility in the form of improved decisions (either now or eventually). Although this idea goes back at least to I. J. Good (1968), it has proved difficult both in theory and in practice. Ultimately, the approach promises to eliminate the notion of algorithm in favour of the notion of adaptive, rational metalevel control of objectlevel computations.