Understanding Reinforcement Learning and its Connection to Computational Neuroscience

Neuroscientists have discovered that networks of neurons can produce different firing patterns based on the weights between the neurons. However, understanding specific biological networks is challenging. Reinforcement learning, which involves learning from rewards rather than feedback, is a difficult problem but has applications in AI. Richard Bellman developed the concept of a value function to determine if a state is good or bad, even without knowing the rewards. The value function can guide decision-making to increase the chances of receiving rewards. Learning the value function involves prediction and comparison of expected and actual rewards. Animal experiments have supported these findings.