Deep Q-Learning in Continuous Control
Prototype
?
?
A simplified 2-D re-creation of the Gymnasium Pusher task: a jointed arm pushes the object (white) onto the goal (red) from randomized spawn cells, trained live in your browser with tabular Q-learning over (fingertip, object) states. Tune the episode budget and reward weights, retrain, and compare runs below — the same reward-convergence analysis documented in the research.
Highlights
- Implemented a standard Deep Q-Network (DQN) reinforcement learning agent from scratch to manipulate coordinate objectives within the OpenAI Gym Pusher-v2 physics engine.
- Trained and updated neural parameters using experience replay memory blocks across 100+ baseline iterations, logging system performance via multi-run moving average reward trajectories.
- Provided comprehensive technical documentation tracking reward convergence dynamics, detailing architectural expansion suggestions including Double-DQN and Dueling-DQN variants to bypass overestimation biases.
Technologies
- Deep Q-Network (DQN)
- OpenAI Gym
- Double-DQN
- Dueling-DQN
- Experience Replay