Analyzing Reinforcement Learning Benchmarks with Random Weight Guessing

TL;DR

This paper introduces a new method using random policy networks to objectively analyze and visualize the complexity of reinforcement learning benchmarks, revealing insights into environment difficulty and benchmark triviality.

Contribution

The study presents Random Weight Guessing as a learning-agnostic approach to evaluate RL benchmarks, providing a baseline and insights into environment complexity without training.

Findings

Random networks can serve as robust baselines.

Some benchmarks are trivial for untrained networks.

The method isolates environment complexity effectively.

Abstract

We propose a novel method for analyzing and visualizing the complexity of standard reinforcement learning (RL) benchmarks based on score distributions. A large number of policy networks are generated by randomly guessing their parameters, and then evaluated on the benchmark task; the study of their aggregated results provide insights into the benchmark complexity. Our method guarantees objectivity of evaluation by sidestepping learning altogether: the policy network parameters are generated using Random Weight Guessing (RWG), making our method agnostic to (i) the classic RL setup, (ii) any learning algorithm, and (iii) hyperparameter tuning. We show that this approach isolates the environment complexity, highlights specific types of challenges, and provides a proper foundation for the statistical analysis of the task's difficulty. We test our approach on a variety of classic control benchmarks from the OpenAI Gym, where we show that small untrained networks can provide a robust baseline for a variety of tasks. The networks generated often show good performance even without gradual learning, incidentally highlighting the triviality of a few popular benchmarks.