Intelligent decision-making through a simulation of evolution

Artificial intelligence can be approached through the fast-time evolution of finite-state machines. Random mutation of an arbitrary machine yields an “offspring.” Both machines are driven by the available history and evaluated in terms of the given goal, and the machine having the higher score is selected to serve as the new parent. Such fast-time mutation and selection is continued with real-time decisions being based on the logic of the surviving machine. Saving the best few machines increases the security against gross nonstationarity of the environment. The efficiency of the evolutionary program is improved by introducing a cost-for-complexity weighting on each machine. An ability to predict one's environment is prerequisite to purposeful behavior. With this in mind, IBM7094 experiments were conducted to examine evolutionary prediction. As expected, cyclic signals in various degrees of noise were soon characterized by the predictor-machines. The transition probabilities within the sequence of predictions of low-order Markov processes were in close correspondence with those of the environment. The evolutionary program was also required to predict the (4-symbol) output sequence of an arbitrary machine that was driven by random binary noise. After 160 predictions the percent correct reached 51.5. When the evolutionary program was also given, the input binary variable this score reached 80 percent, showing a rapid approach toward the 100 percent asymptote. In contrast, providing an uncorrelated binary variable degraded the performance to 40.5 percent by requiring an attempt to extract nonexistent information. A formal technique was devised which translates a predictor machine into a set of hypotheses concerning the logic of the environment.