Genetically programming networks to evolve memory mechanisms
Arlindo Silva, Ana Rute Neves, Ernesto J. F. Costa · 1999
To achieve a high degree of autonomy, an agent usually needs some kind of memory mechanism. In this article we present a new approach to the evolution of agents with memory, based on the use of Genetically Programmed Networks. These are connectionist structures where each node has an associated program, evolved using genetic programming. Genetically Programmed Networks can easily be evolved into agents with very different architectures. We present experimental results from evolving Genetically Programmed Networks as neural networks, distributed programs and rule-based systems capable of solving problems where the use of memory by the agent is essential. Comparisons are made between the performance of these solutions and the performance of solutions obtained by other evolutionary strategies used to evolve agents with memory. 1 -- Introduction Autonomous agents need memory to encode the state information needed to efficiently solve complex problems. In most of the approaches where some...