Metareasoning is reasoning about a reasoning process. This book is a collection of 20 recent research papers by various authors, stemming from a workshop in Chicago in July 2008.
The term “metareasoning” can apply to many different ideas and techniques. In order to “assemble some measure of consistency and soundness in discussing the topic,” the editors wrote a short “manifesto,” laying out a very general, simple model of metareasoning. The authors relate their own work to this framework; they also discuss the relation of their work to the other papers in the collection.
The manifesto discusses three aspects of metareasoning. First, there is a three-level model of an agent: at the ground level, the agent acts in an environment; at the object level, it reasons about its actions and the environment; and at the metalevel, it reasons about its ground-level reasoning. For example, if the agent is using an anytime algorithm to generate a plan, then it must eventually decide that the current plan is acceptable, and that it should act now rather than spend more time thinking; it makes this decision at the metalevel. Second, there is multiagent metareasoning: in a group of interacting agents, each agent reasons about the reasoning processes of the rest. Third, there is self-modeling: the agent’s model of the world should include itself, and, in particular, should include knowledge of its own reasoning abilities. The papers, in fact, mostly lie close to this framework.
I found this book to be unsatisfying on two levels. First, in many of the papers, the description of the object level is inadequate, perhaps because of the 16-page limit. For example, “Using Introspective Reasoning to Improve a CBR System Performance,” by Arcos, Mülâyim, and Leake, describes a case-based reasoning system used in the design of gas treatment plants. The application and approach sound interesting, but the paper offers no description of what the system does, or how it works. One or two concrete examples discussing a specific problem, the comparable cases that were invoked, and the role of metareasoning in picking good cases would be much more meaningful than the result that four different metareasoning strategies reduced the fraction of “low-quality” cases retrieved from 16.67 percent to 14.49 percent, 9.05 percent, 12.85 percent, and 8.21 percent. Many of the other papers have comparable gaps.
The second source of dissatisfaction is that the authors make little attempt to situate the research directions discussed here within the broader context of artificial intelligence (AI) and computer science (CS). For example, several papers deal with some form of scheduling, a subject with an immense amount of CS literature that never invokes metareasoning. Quite possibly, the particular problems here have unusual features that make metareasoning more effective than conventional scheduling techniques, but there is no discussion of what those features are.
Here is another example: Kennedy, in “Distributed Metamanagement,” writes, “[A] robot expects to find an office at the end of a corridor, but instead finds a cupboard. Without metalevel reasoning ... it would just assume that the cupboard is a very small office.” This, however, is just an ordinary problem of belief updating; why is metareasoning more appropriate here than Bayesian updating, belief revision, or nonmonotonic inference?
We can view the same issue from the opposite direction. In his influential article , Kowalski argued that an effective way of modularizing algorithm design is to view the computation as ground inference guided by a control regime. From this standpoint, we can view some substantial fraction of CS (such as dynamic programming, compiler optimization, database query optimization, and adaptive mesh techniques) as manipulating a control regime to carry out inference most effectively. Or consider the selective deepening technique of game tree search, in which particular nodes in a game tree are selected to explore more deeply, a major component of the success of Deep Blue. This conforms precisely to Cox and Raja’s schema: the ground level is the actual execution of chess moves, the object level is the evaluation of the game tree, and the metalevel is the selective deepening decision. None of these is generally described as metareasoning. Should they be? If not, what are the features that divide them from the work described in this book?