Approximating two value functions instead of one: towards characterizing a new family of Deep Reinforcement Learning algorithms
share
This paper makes one step forward towards characterizing a new family of model-free Deep Reinforcement Learning (DRL) algorithms. The aim of these algorithms is to jointly learn an approximation of the state-value function (V), alongside an approximation of the state-action value function (Q). Our analysis starts with a thorough study of the Deep Quality-Value Learning (DQV) algorithm, a DRL algorithm which has been shown to outperform popular techniques such as Deep-Q-Learning (DQN) and Double-Deep-Q-Learning (DDQN) sabatelli2018deep. Intending to investigate why DQV's learning dynamics allow this algorithm to perform so well, we formulate a set of research questions which help us characterize a new family of DRL algorithms. Among our results, we present some specific cases in which DQV's performance can get harmed and introduce a novel off-policy DRL algorithm, called DQV-Max, which can outperform DQV. We then study the behavior of the V and Q functions that are learned by DQV and DQV-Max and show that both algorithms might perform so well on several DRL test-beds because they are less prone to suffer from the overestimation bias of the Q function.
READ FULL TEXT