Observation Space Matters: Benchmark and Optimization Algorithm
arXiv:2011.00756 · doi:10.1109/ICRA48506.2021.9561019
Abstract
Recent advances in deep reinforcement learning (deep RL) enable researchers to solve challenging control problems, from simulated environments to real-world robotic tasks. However, deep RL algorithms are known to be sensitive to the problem formulation, including observation spaces, action spaces, and reward functions. There exist numerous choices for observation spaces but they are often designed solely based on prior knowledge due to the lack of established principles. In this work, we conduct benchmark experiments to verify common design choices for observation spaces, such as Cartesian transformation, binary contact flags, a short history, or global positions. Then we propose a search algorithm to find the optimal observation spaces, which examines various candidate observation spaces and removes unnecessary observation channels with a Dropout-Permutation test. We demonstrate that our algorithm significantly improves learning speed compared to manually designed observation spaces. We also analyze the proposed algorithm by evaluating different hyperparameters.
References in corpus (11)
- Neural Architecture Search with Reinforcement Learning
- Learning agile and dynamic motor skills for legged robots
- DeepMind Control Suite
- Uncertainty-Aware Reinforcement Learning for Collision Avoidance
- Reproducibility of Benchmarked Deep Reinforcement Learning Tasks for Continuous Control
- Implementation Matters in Deep Policy Gradients: A Case Study on PPO and TRPO
- What Matters In On-Policy Reinforcement Learning? A Large-Scale Empirical Study
- Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask
- Learning to Walk in the Real World with Minimal Human Effort
- Learning to Locomote: Understanding How Environment Design Matters for Deep Reinforcement Learning
- Learning to Walk via Deep Reinforcement Learning