Upper Confidence Bound Versus Adaptive Posterior Sampling for Multi-Armed Bandit Game Models

We assess the impact of adopting Algorithmic In-formation Ratio with Adaptive Posterior Sampling (APS) for games that can be modelled as Bernoulli Multi-Armed Bandits. In general, Upper Confidence Bound (UCB) technique is popularly used for sequential decision making tasks because of their better performance. But, our research underlines the advantage of using APS in sequential problems. We emphasize the fact that APS performs better due to its faster convergence and reduced regret compared to UCB. We empirically assess this by implementing various games and comparing the agent's performance using both UCB and APS. Our research advocates that APS is a better choice than U CB for tasks that could be modelled as Bernoulli Multi-Armed Bandits.