Objectives:
- Develop basic interactive learning algorithms: human behavior / demonstrations
- Compare performances of different strategies
In the framework of Reinforcement Learning, here applied to gridworld environments, both optimal and sub-optimal agents can be simulated. Aim of this work is generating such noisily-rational behaviours according to Boltzmann's probability distribution, in order to produce demonstrations that resemble the human way of reasoning.
To run the code, one can choose between:
- a jupiter notebook to be run in the browser -> src/main.ipynb
- a stand-alone python script to be downloaded and run locally -> src/main.py
N.B.: main.py relies on the Grid class, contained in the file src/grid.py
Finally, a report summarising the main results is available in pdf format and some videos showing the agent learning can be found in the videos folder.