Reflexive Interpreters

The goal of achieving powerful problem solving capabilities leads to the “advice taker” form of program and the associated problem of control. This proposal outlines an approach to this problem based on the construction of problem solvers with advanced self-knowledge and introspection capabilities. 1 The Problem of Control Self-reverence, self-knowledge, self-control, These three alone lead life to sovereign power. Alfred, Lord Tennyson, OEnone Know prudent cautious self-control is wisdom’s root. Robert Burns, A Bard’s Epitath The woman that deliberates is lost. Joseph Addison, Cato Amajor goal of Artificial Intelligence is to construct an “advice taker”, a program which can be told new knowledge and advised about how that knowledge may be useful. Many of the approaches towards this goal have proposed constructing additive formalisms for transmitting knowledge to the problem solver. In spite of considerable work along these lines, formalisms for advising problem solvers about the uses and properties of knowledge are relatively undeveloped. As a consequence, the nondeterminism resulting from the uncontrolled application of many independent pieces of knowledge leads to great inefficiencies. Just the potential for nondeterminism requires organizations of problem solvers which are slower on deterministic problems than the corresponding deterministically-tailored systems. The term “advice taker” is from [McCarthy 1968]. The most popular additive formalism is that of the predicate calculus, interpreted by a theorem proving data base. (See [Green 1969], [Darlington 1969], [Fikes and Nilsson 1971], [Moore 1975], and [McDermott 1977a].) The additivity of such a representation stems from the celebrated Tarski [1944] definition of truth for these languages, in which the addition of new truths cannot affect the validity of previous truths. Systems violating this truth-preserving or monotonic property are called non-monotonic systems, and have been studied practically and theoretically by Rescher [1964], Kripke [1975], Doyle [1978], and [McDermott and Doyle 1978]. Several other formalisms including the Carnegie (e.g. [Rychener 1976]) and California (e.g. [Davis 1976]) flavors of production systems and the logical programming languages [Kowalski 1973] are touted as additive, but in practice these rarely are used in an purely additive fashion. Completely deterministic interpreters can be organized so that the execution of each action also sets up the next action to be performed. Interpreters made non-deterministic through additivity must expend much more effort in this regard. Since the set of successor actions is expandable, either considerable work must be done to fetch the relevant successor actions, or considerable work must be done to integrate new actions into the successor action table.