Gradient Temporal-Difference Learning with Regularized Corrections
We demonstrate, for the first time, that Gra- dient TD methods can outperform Q-learning when using neural networks, in two classic control domains and two.
Temporal Difference (Sarsa and Q-Learning)TD methods update their es>mates based in part on other es>mates. They learn a guess from a guess. Is this a good thing to do? Page 21 ... MDP and RL: Q-learning, stochastic approximationTD samples one-step and uses a previous estimation of V . ? DP needs all possible values of V (s?). MC: One full trajectory for update TD: ... Off-Policy Temporal-Difference Learning with Function ApproximationWe introduce the first algorithm for off-policy temporal-difference learning that is stable with linear function approximation. Off- policy learning is of ...
Autres Cours:
