Neural Networks and Reinforcement Learning

A practical approach to neural networks and reinforcement learning.

Demos overview

• Pendulum
  • PD baseline: pendulum/pd_controller.py
  • MLP policy (PD imitation for stability): pendulum/mlp_policy.py
• Double pendulum
  • PD baseline: double_pendulum/pd_controller.py
  • MLP policy (PD imitation): double_pendulum/mlp_policy.py
  • Notebook comparison: double_pendulum/chaos_vs_control.ipynb
• Walker/Hopper
  • Phase-imitation gait demo: walker/hopper_demo.py
  • Reward shaping exploration: walker/reward_shaping.py

Common utilities:

• Plotting & trajectory: common/plotting.py
• Environment wrappers: common/env_wrappers.py (lightweight approximations for reproducible demos)
• Models & IO: common/model_utils.py

Directory Structure

demos/
├── pendulum/
│   ├── pd_controller.py        # Classical PD control demo
│   ├── mlp_policy.py           # Neural network policy demo
│   ├── utils.py                # Shared plotting/logging helpers
│   └── README.md               # Instructions & learning outcomes
│
├── double_pendulum/
│   ├── pd_controller.py        # PD control for double pendulum
│   ├── mlp_policy.py           # Pretrained MLP stabilization demo
│   ├── chaos_vs_control.ipynb  # Notebook comparing PD vs. MLP
│   └── README.md
│
├── walker/
│   ├── hopper_demo.py          # RL-trained hopper gait
│   ├── reward_shaping.py       # Illustrates reward design impact
│   └── README.md
│
├── common/
│   ├── plotting.py             # Graph utilities (angle, torque, reward curves)
│   ├── env_wrappers.py         # PyBullet environment setup helpers
│   └── model_utils.py          # Load/save small MLP models
│
└── README.md                   # Overview of all demos