Self-Adaptation of Meta-Parameters for Lamarckian-Inherited Neuromodulated Neurocontrollers in the Pursuit-Evasion Game

• Published in: 2020 IEEE Symposium Series on Computational Intelligence (SSCI) pp. 2592 - 2599
• Main Authors: Showalter, Ian, Schwartz, Howard
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2020
• Subjects: Evolutionary computation; Evolutionary Meta-parameters; Games; Hebbian Learning; Lamarckian Inheritance; Multi-objective; Neurocontrollers; Neuroevolution; Neuromodulation; Optimization; Parameter Control; Parameter Tuning; Pursuit-Evasion; Sociology; Statistics; Unsupervised Learning
• DOI: 10.1109/SSCI47803.2020.9308450

Determining meta-parameter settings is a longstanding challenge in evolutionary computing, and often involves running repeated simulations with successive estimated values until acceptable values are found. The term metaevolution is often used to describe the optimization of one or more of the meta-parameters that control, for example, the rates of selection and variation in evolutionary optimization algorithms. Here we use self-adaptation to simultaneously evolve all of the practical meta-parameters together with the other evolved parameters as part of the main evolutionary algorithm. The meta-parameters are added to the gene, and evolved along with the other parameters. The evolved neurocontrollers with self-adaptation are compared to those with manually selected meta-parameters. Secondly, self-adapted meta-parameters from the best neurocontroller are used to evolve a further set of non-self-adapted neurocontrollers, and compared with the selfadapted and manually selected results. The effects of selfadaptation are determined through a series of experiments using a previously demonstrated multi-objective Lamarckianinherited neuromodulated evolutionary neurocontroller. The fitness of the evolved neurocontrollers is determined through their operation of a simulated vehicle pursuing a basic evader vehicle in the pursuit-evasion game. Both vehicles are subject to the effects of mass and drag. It is shown that self-adaptation can be used to automatically tune and control meta-parameters during evolution. Only a trivial amount of computational cost is added, and no degradation in evolutionary performance was observed. Under some circumstances self-adaptation may lead to improved performance of the evolutionary algorithm.