Evolving artificial intelligence

Abstract

The majority of research in artificial intelligence has been devoted to modeling the symptoms of intelligent behavior as we observe them in ourselves. Investigation into the causative factors of intelligence have been passed over in order to more rapidly obtain the immediate consequences of intelligence. The results of these efforts have been computer programs that achieve outstanding performance, but only in very limited domains of application. It is suggested that attention be given to the mechanisms that generate intelligence. Intelligence may be defined as that property which enables a system to adapt its behavior to meet desired goals in a range of environments. Three organizational forms of intelligence are characterized within the present discussion: (1) phylogenetic (arising within the phyletic line of descent), (2) ontogenetic (arising within the individual), and (3) sociogenetic (arising within the group). It is argued that all three forms of intelligence are equivalent in process and that all intelligent systems are inherently in nature. Simulating natural evolution provides a method for machine generated intelligent behavior. A series of experiments is conducted to quantify the efficiency and effectiveness of problem solving. The results indicate that this evolutionary programming can rapidly discover nearly optimum solutions to a broad range of problems. Mathematical analysis of the algorithm and its variations indicates that the process will converge to the global best available solution. Automatic control and gaming experiments are conducted in which an program must discover suitable strategies for solving the task at hand. No credit assignment or other heuristic evaluations are offered to the programs. The results indicate the utility of using simulated evolution for general problem solving.