Predictive neural networks for reinforcement learning

By Lukasz Burzawa, Abhishek Chaurasia and Eugenio Culurciello

We used predictive neural network like CortexNet to show that they can speed up reinforcement learning. We used VizDoom rocket basic scenario.

We collected videos of 500 episodes of human game play, and we pre-trained a predictive neural network on those videos.

Speeding up DQN

Predictive networks train faster to play VizDoom scenario when they are pre-trained with human play video and using the DQN algorithm, 1 epoch = 1000 episodes

Speeding up A2C

Predictive networks train faster to play VizDoom scenario when they are pre-trained with human play video and using the A2C algorithm, 1 epoch = 1000 episodes

[Note: y axis is Score, not Time]

About the author

I have almost 20 years of experience in neural networks in both hardware and software (a rare combination). See about me here: Medium, webpage, Scholar, LinkedIn, and more…