The aim of the research presented in this paper is to give a rational agent with bounded resources a basis for selecting and justifying computational actions. The metareasoning in the title should be interpreted as “reasoning about meta-level control policies.”
The author’s general approach to artificial intelligence (AI) is to view it as a constraint optimization problem. Therefore, a normative meta-level theory for the value of computations is the major tool. With such a theory, computations can be regarded as actions, and then standard planning techniques may be in order. This will eventually result in rationally self-governed software architectures.
The major problem is that evaluating computations as actions is hard. One must consider the tradeoffs among the intrinsic utility of the computation, its time cost, space cost, probability of success, probability of revising the result, and so on. These issues can be characterized only in probabilistic terms. Probability and decision theory is thus a main theme of the paper, and much of the metareasoning becomes reasoning about probabilities. Reasoning about first-order probabilities is in general intractable, so ad hoc techniques and heuristics are the only way to do this.
Now, these techniques and heuristics will be reasoned about at some meta-meta-level. This sequence must stop somewhere to avoid a regress problem. That is, in addition to reasoning about the base level and the action level, the meta-meta-level should decide on its own grounds. The main contribution of the paper is a sound mathematical treatment of heuristic decision methods at this meta-meta-level.