- python version 3.9.16
- PyYAML (yaml parser)
- NumPy (high-performance matrix library)
- PyTorch, aka “torch” (tensor & neural network library)
- OpenAI Gym version 0.25.2, aka “gym” (RL environment simulation library)
- Pygame & imageio-ffmpeg (for OpenAI video rendering)
- scikit-learn (sklearn)
- Matplotlib (data visualization)
-
Hw1: Rabbit farm
To make you be familiar with python and its libraries. -
Hw2: Value Function using matrix approach
Numpy is a good tools for solving matrix computing problems. -
Hw3: Value Function & optimal policy using iterative DP method
Getting to know more about RL with "Values" & "Policy". -
Hw4: Monte Carlo with Cat-Mouse environment
Testing environments with Cat-Mouse envs with our First real RL method "Monte Carlo". -
Hw5: Q-Learning with Cat-Mouse env
A "model free" method, computing "Q(S,A)" without building the States as last homework. -
Hw6: Sarsa & Expected Sarsa with epsilon greedy policy
Sarsa is a little bit more conservative strategy with more exploring state while implementing "Epsilon Decay". -
Hw7: Pytorch & CNN
An overview with CNN models & pytorch API. -
Hw8: Semi-gradient TD(0) with NN function approximation
Cat-Mouse env again! Learning our "Values" & "Policy" with NN model. Interesting. -
Hw9: Q-Learning with NN && Acrobot environment
Off-policy Q-Learning sometimes causes "Dead Triad". How do we solve it? -
Hw10: Using Policy-gradient with Monte Carlo to implement the missing code in pg_mc.py
[1] 從根本學習Reinforcement Learning 系列
[2] 强化学习系列(十二):Eligibility Traces