d3rlpy.metrics.scorer.dynamics_reward_prediction_error_scorer¶
-
d3rlpy.metrics.scorer.dynamics_reward_prediction_error_scorer(dynamics, episodes, window_size=1024)[source]¶ Returns MSE of reward prediction (in negative scale).
This metrics suggests how dynamics model is generalized to test sets. If the MSE is large, the dynamics model are overfitting.
\[\mathbb{E}_{s_t, a_t, r_{t+1} \sim D} [(r_{t+1} - r')^2]\]where \(r' \sim T(s_t, a_t)\).
Parameters: - dynamics (d3rlpy.dynamics.base.DynamicsBase) – dynamics model.
- episodes (list(d3rlpy.dataset.Episode)) – list of episodes.
- window_size (int) – mini-batch size to compute.
Returns: negative mean squared error.
Return type: