Hybrid Gradient Temporal-Difference Learning

With GQ(λ), our development of gradient TD methods is almost but not quite complete. There is some evidence that GTD methods are still not as efficient as classical methods on problems where classical methods are applicable. Some of this perhaps cannot be helped. Nevertheless, we would like an algorithm that can learn like classical methods when those methods are sound and like GTD methods for off-policy situations, and we continue to seek such a hybrid algorithm.