The 3D simulation visualises an agent's neural network estimating Q-values for each possible action as it explores a simple environment, alongside its experience replay buffer filling and its target network trailing the main network's weights.
Pick an environment and training speed, then press play to watch the agent explore, sample batches from replay memory, and periodically sync its target network; pause anytime to inspect the current Q-value estimates.
Environment select, training speed slider, play/pause, rebuild
DeepMind's original DQN agent was trained on the exact same network architecture and hyperparameters across all 49 Atari games, with no game-specific tuning, yet still outperformed a professional human tester on the majority of them.
The 3D simulation visualises an agent's neural network estimating Q-values for each possible action as it explores a simple environment, alongside its experience replay buffer filling and its target network trailing the main network's weights.
The 3D simulation visualises an agent's neural network estimating Q-values for each possible action as it explores a simple environment, alongside its experience replay buffer filling and its target network trailing the main network's weights.
Pick an environment and training speed, then press play to watch the agent explore, sample batches from replay memory, and periodically sync its target network; pause anytime to inspect the current Q-value estimates.
DeepMind's original DQN agent was trained on the exact same network architecture and hyperparameters across all 49 Atari games, with no game-specific tuning, yet still outperformed a professional human tester on the majority of them.