How can I evaluate a trained agent in a deterministic way?

[fatcatZF]

How can I evaluate a trained agent in a deterministic way like the evaluate_policy function in stable baselines3?

[Toni-SM]

Hi @fatcatZF

For it, there will be necessary to modify skrl trainer's eval method or agent's act method to get/return the mean_actions in a similar way as discussed in isaac-sim/IsaacLab#1078 (comment)

[Toni-SM]

Feature added in version 1.4.0:

skrl/CHANGELOG.md

Lines 13 to 15 in d57c8ea

	- Define the `stochastic_evaluation` trainer config to allow the use of the actions returned by the agent's model
	as-is instead of deterministic actions (mean-actions in Gaussian-based models) during evaluation.
	Make the return of deterministic actions the default behavior.