Task: Ant-v1
Task: Humanoid-v1
Task: Hopper-v1
Task: HalfCheetah-v1
A number of network architectures are compared according to its losses and similiarity of distribution against the true values. All networks contain a single hidden layer, either fully connected with all inptus/outputs or connected to each output dimension separately and concatenated afterwards.
| Standardized input | Number of Parameters | Loss (Training/Validation) | Similarity | |
|---|---|---|---|---|
| Fully-connected | ✗ | 6,893,585 | .0412/.0369 | .0351 |
| Fully-connected | ✓ | 6,893,585 | .0350/.0385 | .0373 |
| Concatenated | ✓ | 413,457 | .0056/.0069 | .0129 |
| Concatenated | ✓ | 6,615,057 | .0091/.0113 | .0216 |
MIT