[3E5-GS-2] Machine learning

[3E5-GS-2-03] Evolution of metamemory with a neuromodulated neural network in a foraging task

〇Yusuke Takai¹, Reiji Suzuki¹, Takaya Arita¹ (1. Nagoya University Graduate School of Informatics)

Keywords:metamemory, neuroevolution, neural network, evolutionary computation, artificial life

Metamemory is a kind of metacognition, the ability to monitor and control one's own memory. In this study, we attempted a constructive approach to the foraging behavior by forgetting agents in order to clarify the evolutionary scenario and mechanism of more situated metamemory. Specifically, we designed a task in which an agent moves based on the location of feed given at the beginning, while it is better to avoid foraging than to fail foraging due to forgetting the location information. We then conducted evolutionary experiments by implementing neuroevolution, in which learning itself also evolves, using neuromodulation and an evolutionary computing method NEAT. The experimental results showed that we succeeded in evolving agents that tended to choose to avoid more distant food. Analysis of the network revealed a higher-order function of avoidance by referring to a part of the structure that makes location-based foraging decisions, a process that might be equivalent to metamemory.