The In-Sample Softmax for Offline Reinforcement Learning
About
Reinforcement learning (RL) agents can leverage batches of previously collected data to extract a reasonable control policy. An emerging issue in this offline RL setting, however, is that the bootstrapping update underlying many of our methods suffers from insufficient action-coverage: standard max operator may select a maximal action that has not been seen in the dataset. Bootstrapping from these inaccurate values can lead to overestimation and even divergence. There are a growing number of methods that attempt to approximate an \emph{in-sample} max, that only uses actions well-covered by the dataset. We highlight a simple fact: it is more straightforward to approximate an in-sample \emph{softmax} using only actions in the dataset. We show that policy iteration based on the in-sample softmax converges, and that for decreasing temperatures it approaches the in-sample max. We derive an In-Sample Actor-Critic (AC), using this in-sample softmax, and show that it is consistently better or comparable to existing offline RL methods, and is also well-suited to fine-tuning.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Offline Reinforcement Learning | D4RL (various) | HalfCheetah-Medium48.3 | 22 | |
| Locomotion | MuJoCo walker2d medium-replay D4RL | Average Normalized Score69.8 | 16 | |
| Locomotion | MuJoCo halfcheetah (medium) | Normalized Score48.3 | 10 | |
| Locomotion | MuJoCo hopper (medium-replay) | Normalized Score92.1 | 10 | |
| Locomotion | MuJoCo hopper medium | Normalized Score60.3 | 10 | |
| Locomotion | MuJoCo walker2d medium | Normalized Score82.7 | 10 | |
| Locomotion | MuJoCo halfcheetah (medium-replay) | Normalized Score44.3 | 10 | |
| Locomotion | MuJoCo halfcheetah-medium-expert | Normalized Score83.5 | 5 | |
| Locomotion | MuJoCo hopper-medium-expert | Normalized Score93.8 | 5 | |
| Locomotion | MuJoCo walker2d-medium-expert | Normalized Score109 | 5 |