Mitigating Catastrophic Forgetting with Complementary Layered Learning

Catastrophic forgetting is a stability–plasticity imbalance that causes machine learner to lose previously gained knowledge critical for performing task. The occurs in transfer learning, negatively affecting the learner’s performance, particularly neural networks and layered learning. This work proposes complementary learning technique introduces long- short-term memory reduce negative effects of catastrophic forgetting. In particular, this dual system non-neural network approaches evolutionary computation Q-learning instances because these techniques are used develop decision-making capabilities physical robots. Experiments evaluate new augmentation multi-agent simulation, where autonomous unmanned aerial vehicles learn collaborate maneuver survey an area effectively. Through direct-policy value-based experiments, proposed demonstrated significantly improve task performance over standard successfully balancing stability plasticity.