Approximating two value functions instead of one: towards characterizing a new family of Deep Reinforcement Learning algorithms
This work addresses performance issues in deep reinforcement learning for researchers and practitioners, though it is incremental as it builds on prior algorithms like DQV.
The paper tackles the problem of improving deep reinforcement learning by jointly approximating state-value and state-action value functions, showing that their new algorithm DQV-Max outperforms existing methods like DQN and DDQN on several test-beds by reducing overestimation bias.
This paper makes one step forward towards characterizing a new family of \textit{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) \cite{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 \textit{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.