Reinforcement Learning for Tennis Strategy Optimization • Arthur Danjou

Comparison of Reinforcement Learning algorithms on Atari Tennis (ALE/Tennis-v5 via Gymnasium/PettingZoo).

GitHub Repository: Tennis-Atari-Game

Overview

This project implements and compares five RL agents playing Atari Tennis against the built-in AI and in head-to-head tournaments.

Algorithms

Agent	Type	Policy	Update Rule
Random	Baseline	Uniform random	None
SARSA	TD(0), on-policy	ε-greedy	$W_a \leftarrow W_a + \alpha \cdot (r + \gamma \hat{q}(s', a') - \hat{q}(s, a)) \cdot \phi(s)$
Q-Learning	TD(0), off-policy	ε-greedy	$W_a \leftarrow W_a + \alpha \cdot (r + \gamma \max_{a'} \hat{q}(s', a') - \hat{q}(s, a)) \cdot \phi(s)$
Monte Carlo	First-visit MC	ε-greedy	$W_a \leftarrow W_a + \alpha \cdot (G_t - \hat{q}(s, a)) \cdot \phi(s)$
DQN	Deep Q-Network	ε-greedy	MLP (256→256) with experience replay & target network

Architecture

Linear agents (SARSA, Q-Learning, Monte Carlo): $\hat{q}(s, a; \mathbf{W}) = \mathbf{W}_a^\top \phi(s)$ with $\phi(s) \in \mathbb{R}^{128}$ (RAM observation)
DQN: MLP network (128 → 128 → 64 → 18) trained with Adam optimizer, Huber loss, and periodic target network sync

Environment

Game: Atari Tennis via PettingZoo (tennis_v3)
Observation: RAM state (128 features)
Action Space: 18 discrete actions
Agents: 2 players (first_0 and second_0)

Project Structure

.
├── Project_RL_DANJOU_VON-SIEMENS.ipynb    # Main notebook
├── README.md                              # This file
├── checkpoints/                           # Saved agent weights
│   ├── sarsa.pkl
│   ├── q_learning.pkl
│   ├── montecarlo.pkl
│   └── dqn.pkl
└── plots/                                 # Training & evaluation plots
    ├── SARSA_training_curves.png
    ├── Q-Learning_training_curves.png
    ├── MonteCarlo_training_curves.png
    ├── DQN_training_curves.png
    ├── evaluation_results.png
    └── championship_matrix.png

Key Results

Win Rate vs Random Baseline

Agent	Win Rate
SARSA	88.9%
Q-Learning	41.2%
Monte Carlo	47.1%
DQN	6.2%

Championship Tournament

Full round-robin tournament where each agent faces every other agent in both positions (first_0/second_0).

Notebook Sections

Configuration & Checkpoints: incremental training workflow with pickle serialization
Utility Functions: observation normalization and ε-greedy policy
Agent Definitions: RandomAgent, SarsaAgent, QLearningAgent, MonteCarloAgent, DQNAgent
Training Infrastructure: train_agent() and plot_training_curves()
Evaluation: match system, random baseline, and round-robin tournament
Results & Visualization: win rate plots and matchup matrix heatmap

Known Issues

Monte Carlo & DQN: checkpoint loading issues. Saved weights may not restore properly during evaluation, although training works correctly

Dependencies

Python 3.13+
numpy, matplotlib
torch
gymnasium, ale-py
pettingzoo
tqdm

Authors

Arthur DANJOU
Moritz VON SIEMENS