The values below are constructor defaults, not universal recommendations.
| Agent | Learning rate | Discount | Batch/rollout | Main stabilization |
|---|---|---|---|---|
| A2C | 7e-4 |
runner 0.99 |
rollout 5 |
GAE, global gradient clipping |
| DDPG | actor 1e-4, critic 1e-3 |
0.99 |
128 |
target networks, Polyak 0.005 |
| DQN | 1e-3 |
0.99 |
64 |
replay, target update every 500 |
| PPO | 3e-4 |
runner 0.99 |
64, 10 epochs |
ratio clip 0.2, value clipping |
| Q-learning | 0.3 |
0.99 |
transition | tile coding, epsilon 0.05 |
| REINFORCE | 1e-2 |
runner 0.99 |
episode | optional return normalization |
| SAC+gSDE | required arguments | required | required | twin Q, entropy tuning, Polyak |
| True Online Sarsa(λ) | 0.3 |
1.0 |
transition | Dutch traces, lambda 0.9 |
hiddenLayers: dense hidden-layer widths.featureLayers: cloned feature-extractor templates. Each online and target
model receives independent layer instances.maxGradNorm: global gradient norm limit; INF disables clipping.solvedReward: evaluation threshold. null disables early success detection.solvedEvaluations: required consecutive successful evaluations.rewardFunction: training-reward callback.observationFunction: callback converting raw observations to network input.bestModelFile: runner destination for the best evaluated model.Samples commonly read RL_* environment variables. These are sample controls,
not agent API. Inspect the selected sample for its accepted names and defaults.