Simple implementation of meta reinforcement learning experiments in PyTorch.
Dependencies: pytorch 0.4, numpy, matplotlib
Experiments:
-
depbandit_1.ipynb: Dependent bandit experiment I from the Learning to reinforcement learn paper by J. Wang et. al
In this experiment, the agent is trained on a distribution of dependent 2-armed bandits and it is asked to solve particular 2-armed bandits where the reward giving arm is changed in each test episode.
-
depbandit_2.ipynb: Dependent bandit experiment II from the Learning to reinforcement learn paper by J. Wang et. al
In this experiment, the agent is trained on a distribution of dependent 11-armed bandits with deterministic payouts. The nine non-target arms give a reward of 1, the one target arm gives a reward of 5, and the eleventh arm gives an informative reward (<1) of one tenth the target arm's index as 0.2 when the target arm is 2. It is expected from the agent to pay a short-term reward cost to gain information on the target arm and keep pulling it.