Analysis and improvement of policy gradient estimation (情報論的学習理論と機械学習)
スポンサーリンク
概要
- 論文の詳細を見る
Policy gradient is a useful model-free reinforcement learning approach, but it tends to suffer from instability of gradient estimates. In this paper, we analyze and improve the stability of policy gradient methods. We first prove that the variance of gradient estimates in the PGPE (policy gradients with parameter-based exploration) method is smaller than that of the classical REINFORCE method under a mild assumption. We then derive the optimal baseline for PGPE, which contributes to further reducing the variance. We also theoretically show that PGPE with the optimal baseline is more preferable than REINFORCE with the optimal baseline in terms of the variance of gradient estimates. Finally, we demonstrate the usefulness of the improved PGPE method through experiments.
- 社団法人電子情報通信学会の論文
- 2011-06-13
著者
-
Hachiya Hirotaka
Department Of Computer Science Tokyo Institute Of Technology
-
Sugiyama Masashi
Department Of Computer Science Tokyo Institute Of Technology
-
Hachiya Hirotaka
Tokyo Inst. Of Technol.
-
ZHAO Tingting
Department of Computer Science, Tokyo Institute of Technology
-
NIU Gang
Department of Computer Science, Tokyo Institute of Technology
-
Zhao Tingting
Department Of Computer Science Tokyo Institute Of Technology
-
Niu Gang
Department Of Computer Science Tokyo Institute Of Technology
-
Sugiyama Masashi
Department Of Chemistry Faculty Of Science Tokyo University Of Science
-
Sugiyama Masashi
Department of Applied Chemistry, Yamanashi University
関連論文
- Statistical active learning for efficient value function approximation in reinforcement learning (ニューロコンピューティング)
- Improving the Accuracy of Least-Squares Probabilistic Classifiers
- Improving the Accuracy of Least-Squares Probabilistic Classifiers
- Least Absolute Policy Iteration — A Robust Approach to Value Function Approximation
- A New Meta-Criterion for Regularized Subspace Information Criterion
- Approximating the Best Linear Unbiased Estimator of Non-Gaussian Signals with Gaussian Noise
- A new algorithm of non-Gaussian component analysis with radial kernel functions (Special issue: Information geometry and its applications)
- Methods of cross-domain object matching (情報論的学習理論と機械学習)
- Multi-task learning with least-squares probabilistic classifiers (パターン認識・メディア理解)
- Multi-task learning with least-squares probabilistic classifiers (情報論的学習理論と機械学習)
- Adaptive importance sampling with automatic model selection in value function approximation (ニューロコンピューティング)
- Analytic Optimization of Adaptive Ridge Parameters Based on Regularized Subspace Information Criterion(Neural Networks and Bioengineering)
- Adaptive Ridge Learning in Kernel Eigenspace and Its Model Selection
- On Computational Issues of Semi-Supervised Local Fisher Discriminant Analysis
- Recent Advances and Trends in Large-Scale Kernel Methods
- Syntheses of New Artificial Zinc Finger Proteins Containing Trisbipyridine-ruthenium Amino Acid at The N-or C-terminus as Fluorescent Probes
- Improving Model-based Reinforcement Learning with Multitask Learning
- Improving Model-based Reinforcement Learning with Multitask Learning
- Analytic Optimization of Shrinkage Parameters Based on Regularized Subspace Information Criterion(Neural Networks and Bioengineering)
- Least-Squares Conditional Density Estimation
- Constructing Kernel Functions for Binary Regression(Pattern Recognition)
- Optimal design of regularization term and regularization parameter by subspace information criterion
- Information-maximization clustering: analytic solution and model selection (情報論的学習理論と機械学習)
- Conditional Density Estimation Based on Density Ratio Estimation
- Conditional Density Estimation Based on Density Ratio Estimation
- New feature selection method for reinforcement learning: conditional mutual information reveals implicit state-reward dependency (情報論的学習理論と機械学習)
- Least Absolute Policy Iteration-A Robust Approach to Value Function Approximation
- Independent component analysis by direct density-ratio estimation (ニューロコンピューティング)
- A New Meta-Criterion for Regularized Subspace Information Criterion(Pattern Recognition)
- Spectral Methods for Thesaurus Construction
- Adaptive importance sampling with automatic model selection in reward weighted regression (ニューロコンピューティング)
- SERAPH: semi-supervised metric learning paradigm with hyper sparsity (情報論的学習理論と機械学習)
- Analysis and improvement of policy gradient estimation (情報論的学習理論と機械学習)
- Direct density-ratio estimation with dimensionality reduction via hetero-distributional subspace analysis (情報論的学習理論と機械学習)
- Output divergence criterion for active learning in collaborative settings (数理モデル化と問題解決・バイオ情報学)
- Estimation of squared-loss mutual information from paired and unpaired samples (情報論的学習理論と機械学習)
- Dependence minimizing regression with model selection for non-linear causal inference under non-Gaussian noise (情報論的学習理論と機械学習)
- Canonical dependency analysis based on squared-loss mutual information (情報論的学習理論と機械学習)
- Artist agent A[2]: stroke painterly rendering based on reinforcement learning (パターン認識・メディア理解)
- Artist agent A[2]: stroke painterly rendering based on reinforcement learning (情報論的学習理論と機械学習)
- Efficient Sample Reuse in Policy Gradients with Parameter-based Exploration (情報論的学習理論と機械学習)
- Generalization Error Estimation for Non-linear Learning Methods(Neural Networks and Bioengineering)
- Improving Precision of the Subspace Information Criterion(Neural Networks and Bioengineering)
- Canonical dependency analysis based on squared-loss mutual information (パターン認識・メディア理解)
- Change-Point Detection in Time-Series Data by Relative Density-Ratio Estimation (情報論的学習理論と機械学習)
- Modified Newton Approach to Policy Search (情報論的学習理論と機械学習)
- Computationally Efficient Multi-Label Classification by Least-Squares Probabilistic Classifier (情報論的学習理論と機械学習)
- Relative Density-Ratio Estimation for Robust Distribution Comparison (情報論的学習理論と機械学習)
- Change-Point Detection in Time-Series Data by Relative Density-Ratio Estimation
- Modified Newton Approach to Policy Search
- Squared-loss Mutual Information Regularization
- Computationally Efficient Multi-Label Classification by Least-Squares Probabilistic Classifier
- Feature Selection via l_1-Penalized Squared-Loss Mutual Information
- Semi-Supervised Learning of Class Balance under Class-Prior Change by Distribution Matching (情報論的学習理論と機械学習)
- Relative Density-Ratio Estimation for Robust Distribution Comparison
- Artist Agent: A Reinforcement Learning Approach to Automatic Stroke Generation in Oriental Ink Painting
- Winning the Kaggle Algorithmic Trading Challenge with the Composition of Many Models and Feature Engineering (情報論的学習理論と機械学習)
- Computationally Efficient Multi-Label Classification by Least-Squares Probabilistic Classifiers
- Direct Density Ratio Estimation for Large-scale Covariate Shift Adaptation
- Early Stopping Heuristics in Pool-Based Incremental Active Learning for Least-Squares Probabilistic Classifier (情報論的学習理論と機械学習)
- Multi-Task Approach to Reinforcement Learning for Factored-State Markov Decision Problems
- Efficient Sample Reuse in Policy Gradients with Parameter-based Exploration (情報論的学習理論と機械学習)
- Output Divergence Criterion for Active Learning in Collaborative Settings
- Output Divergence Criterion for Active Learning in Collaborative Settings
- Photochromism of benzylviologens containing methyl groups on pyridinium rings and embedded in solid poly(N-vinyl-2-pyrrolidone) matrix.
- Clustering Unclustered Data : Unsupervised Binary Labeling of Two Datasets Having Different Class Balances
- Direct Approximation of Quadratic Mutual Information and Its Application to Dependence-Maximization Clustering
- Direct Learning of Sparse Changes in Markov Networks by Density Ratio Estimation
- Squared-loss Mutual Information Regularization
- Early Stopping Heuristics in Pool-Based Incremental Active Learning for Least-Squares Probabilistic Classifier
- Winning the Kaggle Algorithmic Trading Challenge with the Composition of Many Models and Feature Engineering
- Improving Importance Estimation in Pool-based Batch Active Learning for Approximate Linear Regression