This course taught the foundations of applied reinforcement learning and deep reinforcement learning. I learned to recognize when a problem is a good fit for RL, define it formally with state spaces, action spaces, dynamics, and reward models, and justify which class of algorithms suits it best.
Fundamentals: MDPs and Bandits
The theory track covered Markov decision processes, multi-armed bandits, policy iteration, and dynamic programming. Working through the exploration–exploitation trade-off made concepts like regret and sample complexity concrete, and I learned to distinguish model-based from model-free RL and policy-based from value-based methods.
From Dynamic Programming to TD-Learning
We implemented the classic algorithm families in Python: temporal-difference learning, direct policy search, and function approximation. Evaluating agents properly (convergence behavior, empirical performance, and computational cost) was treated as seriously as training them.
Deep Reinforcement Learning
The deep RL portion connected the fundamentals to state-of-the-art practice: value and policy losses, entropy regularization, and the hyperparameter tuning that makes or breaks training runs. We built working end-to-end pipelines and learned to read training curves rather than trust final scores.
Course Project: RL on Chinese Checkers
I applied the material in reinforcement-learning-on-chinese-checkers, training agents to play Chinese Checkers, a game whose large branching factor and long horizon make reward design and state representation genuinely hard. The project covered the full loop: environment modeling, agent implementation, tuning, and performance evaluation.
After the board-game work I did on Hex with Tsetlin machines, this course rounded out my toolkit for sequential decision making with the gradient-based side of the field.