Intelligent-and-Learning-Agents

Introduction

In a bandit instance, each arm provides a random reward from a probability distribution specific to that arm, this distribution is not known a-priori and it may even change. The objective of the gambler is to maximize the sum of rewards earned through the arms which is same as minimising the ‘regret’.

Tasks Implementation

• Different algorithms for sampling the arms and regret-minimisation of a Bernoulli multi-armed bandit using ε-greedy exploration, UCB, KL-UCB, and Thompson Sampling
• Optimised the scaling coefficient of the exploration bonus in UCB
• An efficient a regret-minimisation algorithm for discrete non Bernoulli bandit instances

Other Details

To run bandit.py add values for the following parameters in command line.

--instance in, where in is a path to the instance file.

--algorithm al, where al is one of epsilon-greedy-t1, ucb-t1, kl-ucb-t1, thompson-sampling-t1, ucb-t2, alg-t3, alg-t4.

--randomSeed rs, where rs is a non-negative integer.

--epsilon ep, where ep is a number in [0, 1]. For everything except epsilon-greedy, pass 0.02.

--scale c, where c is a positive real number.

--threshold th, where th is a number in [0, 1].

--horizon hz, where hz is a non-negative integer.