Towards General-Purpose Model-Free Reinforcement Learning

TL;DR

This paper introduces MR.Q, a model-free deep reinforcement learning algorithm that uses approximate linearization of the value function to perform well across diverse benchmarks with a single hyperparameter set, advancing towards general-purpose RL.

Contribution

The paper proposes MR.Q, a novel model-free RL method leveraging model-based representations for broad applicability without increased complexity or tuning.

Findings

MR.Q performs competitively across multiple RL benchmarks.

It uses a single hyperparameter set for diverse tasks.

The approach bridges the gap between model-based and model-free RL.

Abstract

Reinforcement learning (RL) promises a framework for near-universal problem-solving. In practice however, RL algorithms are often tailored to specific benchmarks, relying on carefully tuned hyperparameters and algorithmic choices. Recently, powerful model-based RL methods have shown impressive general results across benchmarks but come at the cost of increased complexity and slow run times, limiting their broader applicability. In this paper, we attempt to find a unifying model-free deep RL algorithm that can address a diverse class of domains and problem settings. To achieve this, we leverage model-based representations that approximately linearize the value function, taking advantage of the denser task objectives used by model-based RL while avoiding the costs associated with planning or simulated trajectories. We evaluate our algorithm, MR.Q, on a variety of common RL benchmarks with a single set of hyperparameters and show a competitive performance against domain-specific and general baselines, providing a concrete step towards building general-purpose model-free deep RL algorithms.