lambda_return

Contents

lambda_return#

hybrid_aif.train.lambda_return(value, reward, disc, return_lambda)[source]#

DreamerV3’s λ-return over an imagined trajectory.

value is (H+1,), the critic at s_0 … s_H; reward and disc are (H,) indexed by the step taken (reward[k] is collected moving into s_{k+1}, disc[k] = discount · cont(s_{k+1})). Returns (H,), entry k the return estimate at s_k, bootstrapped by value[H].