Back to projects
CompletedRL & RoboticsTeam of 4
Autonomous Taxi Agent
Autonomous navigation agent benchmarking Q-Learning, SARSA, Policy/Value Iteration and DQN.
95 %
DQN success rate
8.5/10
Average reward
< 2 h
Full convergence
Stack
PythonOpenAI GymPyTorchDQNQ-LearningNumPyTensorBoard
- Designing a balanced reward function to guide the learning process
- Optimising hyperparameters for each algorithm tested
- Managing the exploration/exploitation trade-off during training
- Progressive reward system with adaptive penalties
- Grid search and random search for hyperparameter optimisation
- Epsilon-greedy strategy with adaptive decay
- Stabilising DQN training with replay buffer and target network
- Full comparison of several reinforcement learning algorithms (Q-Learning, SARSA, Policy/Value Iteration, DQN)
- DQN achieving a 95% success rate
- Stable convergence in under 2 hours of training
A question about this project?
Contact me