Multi-Arm Bandit Strategies Visualization through a Competitive Space Exploration Game
Live demo here
Watch as two agents using different MAB policies choose between 10 planets to mine for diamonds. Each planet has Gaussian rewards with mean in [0, 100] and variance in [0, 20].
Agents perform an initialization phase where they visit each planet once (or twice for agents following the Bayesian UBC policy in order to get initial variance estimates).
Available MAB policies:
- Random
- Epsilon-Greedy (with epsilon 0.1 and 0.5)
- UCB1
- Bayesian UCB
- EXP3