Optimal Stategies to score runs in cricket using Multi-Armed Bandits

This repository contains several implementations of multi-armed bandit (MAB) agents applied to a simulated cricket batting scenario. The simulation models a cricket innings where an agent (the batsman) selects among different shot strategies (arms) with the goal of maximizing runs while minimizing the risk of getting out.

Overview

In this project, you will find four distinct agent types implemented as part of our exploration of MAB strategies:

1. KL-UCB Survival Agent

   • Description: Uses a KL-divergence based Upper Confidence Bound (UCB) method. The reward is based on survival (i.e., 1 - wicket), focusing on minimizing dismissals.

2. Reward-UCB Agents

   • Variant 1: Reward-UCB (KL) Agent
     • Description: Computes rewards using an efficiency metric (1 - p(out)) * avg_runs and applies a KL-UCB approach.
   • Variant 2: Reward-UCB (Simple) Agent
     • Description: A simpler variant that computes the reward as runs / 6.
   • Variant 3: UCB1 Agent
     • Description: Implements the classic UCB1 algorithm based on the average reward plus an exploration bonus.

3. Risk-Adjusted Successive Elimination Agent

   • Description: Uses a more sophisticated approach by computing a risk-adjusted reward (ratio of expected reward to risk) and eliminates arms that perform poorly. This agent progressively removes suboptimal strategies using confidence bounds.