Speedy Q-Learning
Mohammad Ghavamzadeh, Hilbert J. Kappen, Mohammad Gheshlaghi Azar, Rémi Munos · Neural Information Processing Systems · 2011
We introduce a new convergent variant of Q-learning, called speedy Q-learning (SQL), to address the problem of slow convergence in the standard form of the Q-learning algorithm. We prove a PAC bound on the performance of SQL, which shows that for an MDP with n state-action pairs and the discount factor γ only T = O(log(n)/(e2(1 – γ)4)) steps are required for the SQL algorithm to converge to an ?-optimal action-value function with high probability. This bound has a better dependency on 1/e and 1/(1 2– γ), and thus, is tighter than the best available result for Q-learning. Our bound is also superior to the existing results for both model-free and model-based instances of batch Q-value iteration that are considered to be more efficient than the incremental methods like Q-learning.