Academic project · Reinforcement learning
Multi-armed bandit experiments
How should an agent balance trying new options against using the best one so far? A comparison of classic bandit algorithms on stationary problems.
Approach
Implemented greedy, epsilon-greedy, optimistic initial values and gradient bandit algorithms, and compared their average reward and how often each chose the optimal action across many runs of stationary bandit problems.
Results
Epsilon-greedy with a small exploration rate (epsilon = 0.05) gave the best balance of exploration and exploitation in these experiments.