Lifelong Evolutionary Swarms

Artificial Intelligence (AI) systems are increasingly required to operate autonomously in dynamic environments, adapting to new tasks while retaining previously acquired knowledge—a capability known as lifelong learning. This thesis investigates the integration of lifelong learning into evolutionary swarm intelligence to develop adaptive control strategies for swarm agents. By focusing on evolutionary algorithms, we aim to enable swarms of simple agents to perform complex tasks through collective behavior without central control. We first evaluate various evolutionary algorithms—including NEAT, Genetic Algorithms (GA), Covariance Matrix Adaptation Evolution Strategy (CMA-ES), and EvoStick—in static foraging environments to determine their effectiveness in evolving swarm behaviors. NEAT emerges as the most robust algorithm due to its ability to balance exploration and exploitation effectively. We then modify NEAT to incorporate lifelong learning capabilities, allowing the swarm to adapt to environmental changes or task drifts while maintaining performance. Our experiments demonstrate that evolving a single population through multiple task shifts is as effective as maintaining separate populations for each task, simplifying population management and enhancing computational efficiency. At the population level, neural evolution inherently supports lifelong learning, mitigating catastrophic forgetting as some individuals retain high performance on previous tasks. At the individual level, we observe that regularization techniques based on genetic distance effectively balance the trade-off between retaining past knowledge and learning new tasks, reducing the impact of catastrophic forgetting. In the experiments aimed at learning generalized policies across multiple environments, we observed a drop in performance due to the increased difficulty of the task. However, in some cases, the swarm developed robust behaviors that performed well across diverse scenarios, further validating the effectiveness of the lifelong evolutionary approach. A case study at the Bristol Robotics Laboratory explores both the practical applications and limitations of these evolved controllers in simulated and real-world physical environments. This research contributes to the fields of swarm robotics, evolutionary algorithms, and lifelong learning by providing a framework for developing adaptive, resilient, and autonomous multi-agent systems capable of continuous learning in dynamic environments.