Fast and incremental model-based reinforcement learning

The project focuses on solving the Mountain Car Continuous environment using the Soft Actor-Critic (SAC) as reinforcement learning algorithm. The implementation of SAC was carried out through Stable Baselines3 (SB3) which is a high-level library for reinforcement learning in Python. The primary objective was to train an agent that could efficiently and effectively reach the goal with minimum time and effort.

Throughout the project, we used three model architectures which included Artificial Neural Network (ANN), ANN with Weight Copying, and Recurrent Neural Network (RNN) with Behavioural Cloning. Each model was trained and evaluated using average reward and standard deviation of reward per episode as metrics.

The results proved that both the ANN and RNN with Behavioural Cloning could successfully handle the Mountain Car Continuous environment as they achieved efficient learning and stable performance with medium-sized networks. The project was largely successful upon completion, although minor limitations were observed in the implementation of the ANN with Weight Copying.

Not only did the project achieve its primary objective, but it also provided an opportunity to gain profound insights into reinforcement learning and its practical applications. The experience was invaluable in enhancing our understanding of complex reinforcement learning algorithms and their differences.