ray.rllib.core.rl_module.default_model_config.DefaultModelConfig.free_log_std

DefaultModelConfig.free_log_std: bool = False

If True, for DiagGaussian action distributions (or any other continuous control distribution), make the second half of the policy’s outputs a “free” bias parameter, rather than state-/NN-dependent nodes. In this case, the number of nodes of the policy head have the same dimension as the action space as no slots for log(stddev) are required (only for the mean values).