Adaptive Agents in 1v1 Snake Game with Dynamic Environment

This paper delves into the adaptability of Proximal Policy Optimization (PPO)-trained agents within dynamic environments. Typically, an agent is trained within a specific environment, learning to maximise reward acquisition and to navigate it effectively. However, alterations to this environment can lead to performance deficiencies. Existing research does not fully elucidate how the training of agents influences their adaptability in different environments and which parameters significantly impact this. This study aims to fill this gap, contributing to the creation of more versatile intelligent agents. The objective of this study is to explore how training agents in various environments affects their adaptability when introduced to unfamiliar environments. To this end, 36 models were trained using 36 different configurations to play a one-versus-one (1v1) Snake game. These models were subsequently compared against each configuration to measure their adaptability. The results reveal that map size substantially affect the adaptability of agents in different environments. Interestingly, the results showed that the most adaptive agents were not those trained on the most expansive and complex environment, but rather the simplest.