Astro Teller

Most interesting problems do not have solutions that are simple mappings from the inputs to the correct outputs; some kind of internal state or memory is needed to operate well or optimally in these domains. Traditionally, genetic programming has concentrated on solving problems in the functional/reactive arena. This may be due in part to the absence of a natural way to incorporate memory into the paradigm. This chapter proposes a simple, Turingcomplete addition to the genetic programming paradigm that seamlessly incorporates the evolution of the effective gathering, storage, and retrieval of arbitrarily complicated state information. A new environment is presented and used to evaluate this addition to the paradigm. Experimental results show that the effective production and use of complex memory structures can be evolved and that functions evolving the intelligent use of state quickly and permanently displace purely reactive and non-deterministic functions. These results may aid future research into the causes and constituents of mental models and are shown to open the field of genetic programming to include all learning strategies that are Turing-possible.