d3rlpy.metrics.scorer.average_value_estimation_scorer¶
-
d3rlpy.metrics.scorer.
average_value_estimation_scorer
(algo, episodes, window_size=1024)[source]¶ Returns average value estimation (in negative scale).
This metrics suggests the scale for estimation of Q functions. If average value estimation is too large, the Q functions overestimate action-values, which possibly makes training failed.
\[\mathbb{E}_{s_t \sim D} [ \max_a Q_\theta (s_t, a)]\]Parameters: - algo (d3rlpy.algos.base.AlgoBase) – algorithm.
- episodes (list(d3rlpy.dataset.Episode)) – list of episodes.
- window_size (int) – mini-batch size to compute.
Returns: negative average value estimation.
Return type: