| Implementing Tabular Methods (k3): Students implement fundamental RL concepts, such as the Bellman
Equation, in Python to solve Markov Decision Processes (MDPs) using Policy and Value Iteration (Dynamic
Programing).
Applying Modern RL Libraries (k3): Students are capable of utilizing established high-level libraries (e.g.,
Stable Baselines3, RLlib, CleanRL) to configure and train a wide range of state-of-the-art algorithms (such as
PPO, SAC, TD3).
Solving Continuous Control Problems (k4): Students can select suitable algorithms for continuous control
tasks and apply them to modern simulation environments (e.g., Gymnasium, MuJoCo, PyBullet).
Analyzing and Comparing Algorithms (k5): Students can evaluate the suitability of different algorithms (On-
Policy vs. Off-Policy) for specific tasks and compare their learning progress using monitoring tools (e.g.,
TensorBoard, Weights & Biases).
Hyperparameter Tuning (k5): Students understand the sensitivity of modern algorithms to hyperparameters
(such as learning rate, batch size, entropy coefficient) and can systematically optimize them to obtain stable
policies using modern tools (like Optuna).
|
Students know the mathematical foundations of MDPs, returns, and value functions, as well as the
distinction between model-based vs. model-free, on-policy vs. off-policy approaches. They have a comprehensive
overview of the landscape of modern Deep RL algorithms (such as DQN, PPO, SAC, TD3) and know which
algorithm is suitable for which type of problem (discrete vs. continuous, stochastic vs. deterministic). Furthermore,
they are familiar with the functioning of standard interfaces for RL environments.
|
