Training and evaluation environments are separate constructor arguments. Evaluation uses deterministic policy methods and does not add epsilon, Gaussian, or OU exploration. SAC may optionally evaluate with gSDE noise.
evaluate() returns mean raw reward. Runners exposing evaluateDetailed()
return:
[
'rawReward' => float,
'transformedReward' => float,
'steps' => float,
]
Raw reward determines task success. Transformed reward describes the training signal and can distinguish checkpoints tied on raw score. An episode ends on either termination or truncation.
When bestModelFile is set, the runner saves only improvements. solvedReward
and solvedEvaluations control early completion independently of checkpoint
saving.