Evolution of Intelligent Agents

Most of the work since submitting the previous report has concentrated on software implementation. We have developed a platform in which artificial creatures composed of rigid three-dimensional (3D) blocks and controlled by neural networks may interact physically with their environment and with each other. The design of this platform is mostly based on the experiments of Karl Sims [6, 5], with important differences. We succeded in building a stable platform in which efficient behaviours consistently emerge in a reasonable time (successful locomotion evolves within hours). Using this platform, we have performed a series of experiments in evolution and co-evolution, again similar to Sims’. We have described our initial experiments and their results in a paper which we submitted to the GECCO conference. The paper was accepted for poster presentation, but we chose not to present at GECCO and instead to try and improve the paper in order to obtain a full oral presentation at the CEC conference. Much of this report is taken from the paper submitted to the CEC conference. In the following sections we provide a broad description of our system, stressing both similarities and differences with Sims’ model. In order to facilitate comparisons, our description deliberately follows the same organisation as Sims [6], section by section. At the end of this paper, we provide a table which summarises the important numerical constants used for our experiments (see Table 1).