Dopamine in the Mitigation of Catastrophic Forgetting in Reinforcement Learning

Catastrophic forgetting presents a major challenge in reinforcement learning. It is the situation where a model forgets previously learned tasks while learning new ones. Retention of long-term knowledge for the model becomes difficult while adapting to new environments. This paper explores how dopamine-driven learning mechanisms in the human brain can be integrated into reinforcement learning to address this issue. Specifically, we experiment with dopamine-inspired modifications to Deep Q-Networks (DQN) by leveraging temporal difference learning in novel ways. The approaches include adjusting the learning rate through Neuromodulated Plasticity, incorporating dopamine-driven signals as a loss function for temporal difference (TD) learning, hebbian learning with eligibility traits and dopamine-gated synaptic consolidation (DNC). We compare these methods with standard DQN and DQN with Elastic Weight Consolidation (EWC) to see how they affect the learning efficiency, adaptability, and long-term retention. Furthermore, we analyze the role of these biologically inspired mechanisms in mitigating catastrophic forgetting, aiming to enhance the stability of learned policies across sequential tasks.