← back

Policy gradient on GridWorld

In this experiment, we play around with a very simple algorithm on a very simple environment. This is a good starting point for RL research: train a neural net with the policy gradient method, experimenting with different experimental setups, such as varying the size of the model, training episode length, and the size of the environment/grid.

Base: 1 layer, random policy

• 5000 episodes
• 1 layer MLP
• GridWorld size: 10
• Training algorithm: none (random)

Base: 1 layer, trained policy

• 5000 episodes
• 1 layer MLP
• GridWorld size: 10
• Training algorithm: REINFORCE with policy gradient