An artificial intelligence technique to generate self-optimizing experimental designs

Abstract

The paper describes a completed and independent module of a large-scale system, the Quasi-Optimizer (<u>QO</u>). The <u>QO</u> system has three major objectives: (i) to observe and measure adversaries' behavior in a competitive environment, to infer their strategies and to construct a computer model, a <u>descriptive theory</u>, of each; (ii) to identify strategy components, evaluate their effectiveness and to select the most satisfactory ones from a set of computed descriptive theories; (iv) to combine these components in a quasi-optimum strategy that represents a <u>normative theory</u> in the statistical sense.The measurements on the input strategies can take place either in a sequence of confrontations unperturbed by the <u>QO</u>, or, for efficiency's sake, in a series of environments specified according to some experimental design. The module completed first, <u>QO-I</u>, can perform the experiments either in an exhaustive manner---when every level of a decision variable is combined with every level of the other decision variables---or, in relying on the assumption of a monotonically changing response surface, it uses the binary chopping technique.The module discussed here, <u>QO-3</u>, does not assume monotonic response surfaces and can deal also with multidimensional responses. It starts with a (loosely) balanced incomplete block design for the experiments and computes dynamically the specifications for each subsequent experiment. Accordingly, the levels of the decision variables in any single experiment and the length of the whole sequence of experiments depend on the responses obtained in previous experiments. In general, <u>QO-3</u> is an on-line, dynamic generator of experimental design that minimizes the total number of experiments performed for a predetermined level of precision.