SUNA-X: Using Novelty Retention to Improve Spectrum-diverse Neuroevolution with Unified Neural Models in Complex Environments

Sammanfattning

Adaptability is essential for agents to learn in dynamic and unpredictable environments in artificial intelligence. We explore how to enhance Spectrum-diverse Neuroevolution with Unified Neural Models (SUNA) by integrating Novelty Map (Nmap) to improve its adaptability, this updated version we call SUNA-X. SUNA is an evolutionary algorithm that can develop diverse neural network topologies but faces challenges in maintaining adaptability under changing conditions. The integration of Nmap addresses this limitation by retaining and utilizing novel experiences, enhancing the agent’s capacity to adapt to new and unforeseen situations. To evaluate the effectiveness of this approach, we developed a complex training environment named EcoSim-AI using Godot Engine. EcoSim-AI simulates an ecological system where multiple agents with arbitrary objectives interact and learn within a shared environment. This allows SUNA-X adaptability to be tested. The results show that agents using SUNA-X achieve higher rewards and exhibit more robust adaptability than those using SUNA.