[Paper Review] Benchmarking Model-Based Reinforcement Learning

Tingwu Wang, Xuchan Bao+8 more|arXiv (Cornell University)|Jul 3, 2019

Reinforcement Learning in Robotics43 references242 citations

TL;DR

A comprehensive benchmark of 11 MBRL algorithms and 4 MFRL baselines across 18 OpenAI Gym-like environments, analyzing performance, robustness to noise, and three bottlenecks in model-based RL.

ABSTRACT

Model-based reinforcement learning (MBRL) is widely seen as having the potential to be significantly more sample efficient than model-free RL. However, research in model-based RL has not been very standardized. It is fairly common for authors to experiment with self-designed environments, and there are several separate lines of research, which are sometimes closed-sourced or not reproducible. Accordingly, it is an open question how these various existing MBRL algorithms perform relative to each other. To facilitate research in MBRL, in this paper we gather a wide collection of MBRL algorithms and propose over 18 benchmarking environments specially designed for MBRL. We benchmark these algorithms with unified problem settings, including noisy environments. Beyond cataloguing performance, we explore and unify the underlying algorithmic differences across MBRL algorithms. We characterize three key research challenges for future MBRL research: the dynamics bottleneck, the planning horizon dilemma, and the early-termination dilemma. Finally, to maximally facilitate future research on MBRL, we open-source our benchmark in http://www.cs.toronto.edu/~tingwuwang/mbrl.html.

Motivation & Objective

• Assess relative performance of contemporary MBRL algorithms against model-free baselines under standardized, shared benchmarks.
• Evaluate robustness of MBRL methods to observation and action noise.
• Identify and characterize core challenges limiting MBRL progress (dynamics bottleneck, planning horizon, early termination).
• Provide open-source benchmarking resources to enable reproducibility and fair comparisons.

Proposed method

• Compile a diverse set of 11 MBRL algorithms and 4 MFRL baselines.
• Standardize environments (18 tasks) and problem settings (including noise) based on OpenAI Gym; modify rewards to ensure differentiability for certain methods.
• Evaluate performance at 200k time-steps (and 1M time-steps for select methods) with four random seeds, using grid-searched hyperparameters per algorithm.
• Analyze robustness to observation and action noise via Gaussian perturbations.
• Investigate three hypothesis-driven bottlenecks by empirical measurement (dynamics bottleneck, planning horizon, early termination).
• Provide an open-source benchmark platform and documentation for reproducibility.

Experimental results

Research questions

• RQ1 How do existing MBRL approaches compare to each other and to standard MFRL baselines across a spectrum of environment difficulties?
• RQ2 Are MBRL methods robust to observation and action noise, and how does this robustness compare to model-free baselines?
• RQ3 What are the main factors limiting MBRL performance, and do they manifest as dynamics bottlenecks, planning horizon issues, or early termination dilemmas?
• RQ4 Can standardized benchmarks and open-source code accelerate progress and reproducibility in MBRL?

Key findings

• No single MBRL method dominates across all environments; performance varies with task difficulty and environment characteristics.
• Shooting and Dyna-style MBRL methods often excel on simpler tasks, while complex, high-dimensional tasks reveal remaining gaps compared to model-free methods and ground-truth dynamics.
• Robustness to observation and action noise is heterogeneous; some Dyna-style methods show resilience, whereas others degrade more with noise.
• Three persistent bottlenecks are identified: dynamics bottleneck (learned dynamics plateauing with more data), planning horizon dilemma (long horizons can hurt performance due to curse of dimensionality and model errors), and early termination dilemma (early termination often harms MBRL performance).
• Ground-truth dynamics generally enable higher performance, but cannot always be scaled; when dynamics are learned, performance plateaus at levels below model-free baselines and full-ground-truth baselines in several tasks.
• The study emphasizes the importance of uncertainty modeling, ensembles, and robust planning modules to mitigate model bias and extrapolation errors.