Loaded DiCE: Trading off Bias and Variance in Any-Order Score Function Estimators for Reinforcement Learning

2019/09/23 by Gregory Farquhar, Shimon Whiteson, Farquhar, Gregory +3 · 1 citation
Computer Science · Mathematics · Psychology · #Advanced Multi-Objective Optimization Algorithms #Data Stream Mining Techniques #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Mental Health Research Topics #Reinforcement Learning in Robotics #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1909.10549

arxiv created 2019/09/23 · openalex publication_date 2019/09/23 · arxiv updated 2019/09/25 · openalex created_date 2022/09/13 · openalex updated_date 2026/07/28

Abstract

Gradient-based methods for optimisation of objectives in stochastic settings with unknown or intractable dynamics require estimators of derivatives. We derive an objective that, under automatic differentiation, produces low-variance unbiased estimators of derivatives at any order. Our objective is compatible with arbitrary advantage estimators, which allows the control of the bias and variance of any-order derivatives when using function approximation. Furthermore, we propose a method to trade off bias and variance of higher order derivatives by discounting the impact of more distant causal dependencies. We demonstrate the correctness and utility of our objective in analytically tractable MDPs and in meta-reinforcement-learning for continuous control.

Citations

• High-Dimensional Continuous Control Using Generalized Advantage\n Estimation
• Gradient Estimation Using Stochastic Computation Graphs
• The Optimal Reward Baseline for Gradient-Based Reinforcement Learning
• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• Some Considerations on Learning to Explore via Meta-Reinforcement Learning
• Counterfactual Multi-Agent Policy Gradients
• An Introduction to Deep Reinforcement Learning

Cited by

• Robust and Diverse Multi-Agent Learning via Rational Policy Gradient

Related

• Unifying Gradient Estimators for Meta-Reinforcement Learning via Off-Policy Evaluation
• DiCE: The Infinitely Differentiable Monte-Carlo Estimator
• Total stochastic gradient algorithms and applications in reinforcement learning
• An Empirical Analysis of Measure-Valued Derivatives for Policy Gradients
• An Analysis of Measure-Valued Derivatives for Policy Gradients