Thompson Sampling for Gridworld

Final project for Reinforcement Learning course

Authors: Lorenzo Basile, Irene Brugnara

Source files:

• gridworld.py contains the implementation of the class Gridworld, defining a two-dimensional grid environment in which an agent moves from an initial cell to a target cell. The position of the target cell is not known, but at each time step the agent receives a random binary signal from the target depending on its distance from the target. The method gridworld_search implements a search algorithm based on Thompson sampling (or a greedy algorithm if greedy=True);
• animation.py contains an example animated run of the search algorithm;
• benchmark.py and analysis.py respectively contain code to collect and process data on larger-scale runs of both Thompson algorithm and greedy algorithm (the former produces a pickle file which is to be read by the latter).