DOCSLIB.ORG

Revisiting the Softmax Bellman Operator: New Benefits and New Perspective

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Revisiting the Softmax Bellman Operator: New Benefits and New Perspective Revisiting the Softmax Bellman Operator: New Benefits and New Perspective Zhao Song 1 * Ronald E. Parr 1 Lawrence Carin 1 Abstract tivates the use of exploratory and potentially sub-optimal actions during learning, and one commonly-used strategy The impact of softmax on the value function itself is to add randomness by replacing the max function with in reinforcement learning (RL) is often viewed as the softmax function, as in Boltzmann exploration (Sutton problematic because it leads to sub-optimal value & Barto, 1998). Furthermore, the softmax function is a (or Q) functions and interferes with the contrac- differentiable approximation to the max function, and hence tion properties of the Bellman operator. Surpris- can facilitate analysis (Reverdy & Leonard, 2016). ingly, despite these concerns, and independent of its effect on exploration, the softmax Bellman The beneficial properties of the softmax Bellman opera- operator when combined with Deep Q-learning, tor are in contrast to its potentially negative effect on the leads to Q-functions with superior policies in prac- accuracy of the resulting value or Q-functions. For exam- tice, even outperforming its double Q-learning ple, it has been demonstrated that the softmax Bellman counterpart. To better understand how and why operator is not a contraction, for certain temperature pa- this occurs, we revisit theoretical properties of the rameters (Littman, 1996, Page 205). Given this, one might softmax Bellman operator, and prove that (i) it expect that the convenient properties of the softmax Bell- converges to the standard Bellman operator expo- man operator would come at the expense of the accuracy nentially fast in the inverse temperature parameter, of the resulting value or Q-functions, or the quality of the and (ii) the distance of its Q function from the resulting policies. In this paper, we demonstrate that, in optimal one can be bounded. These alone do not the case of deep Q-learning, this expectation is surprisingly explain its superior performance, so we also show incorrect. We combine the softmax Bellman operator with that the softmax operator can reduce the over- the deep Q-network (DQN)(Mnih et al., 2015) and double estimation error, which may give some insight DQN (DDQN)(van Hasselt et al., 2016a) algorithms, by re- into why a sub-optimal operator leads to better placing the max function therein with the softmax function, performance in the presence of value function in the target network. We then test the variants on several approximation. A comparison among different games in the Arcade Learning Environment (ALE)(Belle- Bellman operators is then presented, showing the mare et al., 2013), a standard large-scale deep RL testbed. trade-offs when selecting them. The results show that the variants using the softmax Bell- man operator can achieve higher test scores, and reduce the 1. Introduction Q-value overestimation as well as the gradient noise on most of them. This effect is independent of exploration and is The Bellman equation (Bellman, 1957) has been a funda- entirely attributable to the change in the Bellman operator. arXiv:1812.00456v2 [cs.LG] 19 May 2019 mental tool in reinforcement learning (RL), as it provides a This surprising result suggests that a deeper understanding sufficient condition for the optimal policy in dynamic pro- of the softmax Bellman operator is warranted. To this end, gramming. The use of the max function in the Bellman we prove that starting from the same initial Q-values, we can equation further suggests that the optimal policy should be upper and lower bound how far the Q-functions computed greedy w.r.t. the Q-values. On the other hand, the trade-off with the softmax operator can deviate from those computed between exploration and exploitation (Thrun, 1992) mo- with the regular Bellman operator. We further show that the 1Duke University ∗Work performed as a graduate student softmax Bellman operator converges to the optimal Bellman at Duke University; now at Baidu Research. Correspondence operator in an exponential rate w.r.t. the inverse temperature to: Zhao Song <[email protected]>, Ronald E. Parr parameter. This gives insight into why the negative conver- <[email protected]>, Lawrence Carin <[email protected]>. gence results may not be as discouraging as they initially Proceedings of the 36 th International Conference on Machine seem, but it does not explain the superior performance ob- Learning, Long Beach, California, PMLR 97, 2019. Copyright served in practice. Motivated by recent work (van Hasselt 2019 by the author(s). et al., 2016a; Anschel et al., 2017) targeting bias and insta- Revisiting the Softmax Bellman Operator: New Benefits and New Perspective bility of the original DQN (Mnih et al., 2015), we further a probability mass function (PMF), where π(s; a) 2 [0; 1] investigate whether the softmax Bellman operator can alle- denotes the probability of selecting action a in state s, and P viate these issues. As discussed in van Hasselt et al.(2016a), a2A π(s; a) = 1. one possible explanation for the poor performance of the For a given policy π, its state-action value func- vanilla DQN on some Atari games was the overestimation tion Qπ(s; a) is defined as the accumulated, expected, bias when computing the target network, due to the max discount reward, when taking action a in state s, operator therein. We prove that given the same assumptions and following policy π afterwards, i.e., Qπ(s; a) = as van Hasselt et al.(2016a), the softmax Bellman operator [P1 γtr js = s; a = a] : For the optimal policy can reduce the overestimation bias, for any inverse tem- Eat∼π t=0 t 0 0 π∗, its corresponding Q-function satisfies the following Bell- perature parameters. We also quantify the overestimation man equation: reduction by providing its lower and upper bounds. X Q∗(s; a) = R(s; a) + γ P (s0js; a) max Q∗(s0; a0): Our results are complementary to and add new motivations a0 for existing work that explores various ways of softening s0 the Bellman operator. For example, entropy regularizers In DQN (Mnih et al., 2015), the Q-function is parameterized have been used to smooth policies. The motivations for with a neural network as Qθ(s; a), which takes the state s such approaches include computational convenience, ex- as input and outputs the corresponding Q-value in the final ploration, or robustness (Fox et al., 2016; Haarnoja et al., fully-connected linear layer, for every action a. The training 2017; Schulman et al., 2017; Neu et al., 2017). With respect objective for the DQN can be represented as 2 to value functions, Asadi & Littman(2017) proposed an 1 0 0 min Qθ(s; a) − [R(s; a) + γ max Qθ− (s ; a )] ; alternative mellowmax operator and proved that it is a con- θ 2 a0 traction. Their experimental results suggested that it can (1) improve exploration, but the possibility that the sub-optimal where θ− corresponds to the frozen weights in the target Bellman operator could, independent of exploration, lead to network, and is updated at fixed intervals. The optimization superior policies was not considered. Our results, therefore, of Eq. (1) is performed via RMSProp (Tieleman & Hinton, provide additional motivation for further study of operators 2012), with mini-batches sampled from a replay buffer. such as mellowmax. Although very similar to softmax, the To reduce the overestimation bias, the double DQN mellowmax operator needs some extra computation to rep- (DDQN) algorithm modified the target that Q (s; a) aims resent a policy, as noted in Asadi & Littman(2017). Our θ to fit in Eq. (1) as paper discusses this and other trade-offs among different 0 0 − Bellman operators. R(s; a) + γ Qθ s ; arg max Qθt (s ; a) : a The rest of this paper is organized as follows: We pro- Note that a separate network based on the latest estimate θt vide the necessary background and notation in Section2. is employed for action selection, and the evaluation of this The softmax Bellman operator is introduced in Section3, policy is due to the frozen network. where its convergence properties and performance bound are provided. Despite being sub-optimal on Q-functions, the Notation The softmax function is defined as softmax operator is shown in Section4 to consistently out- [exp(τx ); exp(τx );:::; exp(τx )]T perform its max counterpart on several Atari games. Such 1 2 m fτ (x) = Pm ; surprising result further motivates us to investigate why this i=1 exp(τxi) happens in Section5. A thorough comparison among differ- where the superscript T denotes the vector transpose. Sub- ent Bellman operators is presented in Section6. Section7 sequently, the softmax-weighted function is represented as T discusses the related work and Section8 concludes this gx(τ) = fτ (x) x, as a function of τ. Also, we define the T paper. vector Q(s; ) = [Q(s; a1);Q(s; a2);:::;Q(s; am)] . We further set m to be the size of the action set A in the MDP. 2. Background and Notation Finally, Rmin and Rmax denote the minimum and the maxi- mum immediate rewards, respectively. A Markov decision process (MDP) can be represented as a 5-tuple hS; A; P; R; γi, where S is the state space, A is 3. The Softmax Bellman Operator the action space, P is the transition kernel whose element P (s0js; a) denotes the transition probability from state s We start by providing the following standard Bellman oper- to state s0 under action a, R is a reward function whose ator: element R(s; a) denotes the expected reward for executing X T Q(s; a) = R(s; a) + γ P (s0js; a) max Q(s0; a0): action a in state s, and γ 2 (0; 1) is the discount factor.
Load more

Recommended publications
  • Lecture 4 Feedforward Neural Networks, Backpropagation
    CS7015 (Deep Learning): Lecture 4 Feedforward Neural Networks, Backpropagation Mitesh M. Khapra Department of Computer Science and Engineering Indian Institute of Technology Madras 1/9 Mitesh M. Khapra CS7015 (Deep Learning): Lecture 4 References/Acknowledgments See the excellent videos by Hugo Larochelle on Backpropagation 2/9 Mitesh M. Khapra CS7015 (Deep Learning): Lecture 4 Module 4.1: Feedforward Neural Networks (a.k.a. multilayered network of neurons) 3/9 Mitesh M. Khapra CS7015 (Deep Learning): Lecture 4 The input to the network is an n-dimensional hL =y ^ = f(x) vector The network contains L − 1 hidden layers (2, in a3 this case) having n neurons each W3 b Finally, there is one output layer containing k h 3 2 neurons (say, corresponding to k classes) Each neuron in the hidden layer and output layer a2 can be split into two parts : pre-activation and W 2 b2 activation (ai and hi are vectors) h1 The input layer can be called the 0-th layer and the output layer can be called the (L)-th layer a1 W 2 n×n and b 2 n are the weight and bias W i R i R 1 b1 between layers i − 1 and i (0 < i < L) W 2 n×k and b 2 k are the weight and bias x1 x2 xn L R L R between the last hidden layer and the output layer (L = 3 in this case) 4/9 Mitesh M. Khapra CS7015 (Deep Learning): Lecture 4 hL =y ^ = f(x) The pre-activation at layer i is given by ai(x) = bi + Wihi−1(x) a3 W3 b3 The activation at layer i is given by h2 hi(x) = g(ai(x)) a2 W where g is called the activation function (for 2 b2 h1 example, logistic, tanh, linear, etc.) The activation at the output layer is given by a1 f(x) = h (x) = O(a (x)) W L L 1 b1 where O is the output activation function (for x1 x2 xn example, softmax, linear, etc.) To simplify notation we will refer to ai(x) as ai and hi(x) as hi 5/9 Mitesh M.
    [Show full text]
  • On the Learning Property of Logistic and Softmax Losses for Deep Neural Networks
    The Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI-20) On the Learning Property of Logistic and Softmax Losses for Deep Neural Networks Xiangrui Li, Xin Li, Deng Pan, Dongxiao Zhu∗ Department of Computer Science Wayne State University {xiangruili, xinlee, pan.deng, dzhu}@wayne.edu Abstract (unweighted) loss, resulting in performance degradation Deep convolutional neural networks (CNNs) trained with lo- for minority classes. To remedy this issue, the class-wise gistic and softmax losses have made significant advancement reweighted loss is often used to emphasize the minority in visual recognition tasks in computer vision. When training classes that can boost the predictive performance without data exhibit class imbalances, the class-wise reweighted ver- introducing much additional difficulty in model training sion of logistic and softmax losses are often used to boost per- (Cui et al. 2019; Huang et al. 2016; Mahajan et al. 2018; formance of the unweighted version. In this paper, motivated Wang, Ramanan, and Hebert 2017). A typical choice of to explain the reweighting mechanism, we explicate the learn- weights for each class is the inverse-class frequency. ing property of those two loss functions by analyzing the nec- essary condition (e.g., gradient equals to zero) after training A natural question then to ask is what roles are those CNNs to converge to a local minimum. The analysis imme- class-wise weights playing in CNN training using LGL diately provides us explanations for understanding (1) quan- or SML that lead to performance gain? Intuitively, those titative effects of the class-wise reweighting mechanism: de- weights make tradeoffs on the predictive performance terministic effectiveness for binary classification using logis- among different classes.
    [Show full text]
  • CS281B/Stat241b. Statistical Learning Theory. Lecture 7. Peter Bartlett
    CS281B/Stat241B. Statistical Learning Theory. Lecture 7. Peter Bartlett Review: ERM and uniform laws of large numbers • 1. Rademacher complexity 2. Tools for bounding Rademacher complexity Growth function, VC-dimension, Sauer’s Lemma − Structural results − Neural network examples: linear threshold units • Other nonlinearities? • Geometric methods • 1 ERM and uniform laws of large numbers Empirical risk minimization: Choose fn F to minimize Rˆ(f). ∈ How does R(fn) behave? ∗ For f = arg minf∈F R(f), ∗ ∗ ∗ ∗ R(fn) R(f )= R(fn) Rˆ(fn) + Rˆ(fn) Rˆ(f ) + Rˆ(f ) R(f ) − − − − ∗ ULLN for F ≤ 0 for ERM LLN for f |sup R{z(f) Rˆ}(f)| + O(1{z/√n).} | {z } ≤ f∈F − 2 Uniform laws and Rademacher complexity Definition: The Rademacher complexity of F is E Rn F , k k where the empirical process Rn is defined as n 1 R (f)= ǫ f(X ), n n i i i=1 X and the ǫ1,...,ǫn are Rademacher random variables: i.i.d. uni- form on 1 . {± } 3 Uniform laws and Rademacher complexity Theorem: For any F [0, 1]X , ⊂ 1 E Rn F O 1/n E P Pn F 2E Rn F , 2 k k − ≤ k − k ≤ k k p and, with probability at least 1 2exp( 2ǫ2n), − − E P Pn F ǫ P Pn F E P Pn F + ǫ. k − k − ≤ k − k ≤ k − k Thus, P Pn F E Rn F , and k − k ≈ k k R(fn) inf R(f)= O (E Rn F ) . − f∈F k k 4 Tools for controlling Rademacher complexity 1.
    [Show full text]
  • Loss Function Search for Face Recognition
    Loss Function Search for Face Recognition Xiaobo Wang * 1 Shuo Wang * 1 Cheng Chi 2 Shifeng Zhang 2 Tao Mei 1 Abstract Generally, the CNNs are equipped with classification loss In face recognition, designing margin-based (e.g., functions (Liu et al., 2017; Wang et al., 2018f;e; 2019a; Yao angular, additive, additive angular margins) soft- et al., 2018; 2017; Guo et al., 2020), metric learning loss max loss functions plays an important role in functions (Sun et al., 2014; Schroff et al., 2015) or both learning discriminative features. However, these (Sun et al., 2015; Wen et al., 2016; Zheng et al., 2018b). hand-crafted heuristic methods are sub-optimal Metric learning loss functions such as contrastive loss (Sun because they require much effort to explore the et al., 2014) or triplet loss (Schroff et al., 2015) usually large design space. Recently, an AutoML for loss suffer from high computational cost. To avoid this problem, function search method AM-LFS has been de- they require well-designed sample mining strategies. So rived, which leverages reinforcement learning to the performance is very sensitive to these strategies. In- search loss functions during the training process. creasingly more researchers shift their attention to construct But its search space is complex and unstable that deep face recognition models by re-designing the classical hindering its superiority. In this paper, we first an- classification loss functions. alyze that the key to enhance the feature discrim- Intuitively, face features are discriminative if their intra- ination is actually how to reduce the softmax class compactness and inter-class separability are well max- probability.
    [Show full text]
  • Deep Neural Networks for Choice Analysis: Architecture Design with Alternative-Specific Utility Functions Shenhao Wang Baichuan
    Deep Neural Networks for Choice Analysis: Architecture Design with Alternative-Specific Utility Functions Shenhao Wang Baichuan Mo Jinhua Zhao Massachusetts Institute of Technology Abstract Whereas deep neural network (DNN) is increasingly applied to choice analysis, it is challenging to reconcile domain-specific behavioral knowledge with generic-purpose DNN, to improve DNN's interpretability and predictive power, and to identify effective regularization methods for specific tasks. To address these challenges, this study demonstrates the use of behavioral knowledge for designing a particular DNN architecture with alternative-specific utility functions (ASU-DNN) and thereby improving both the predictive power and interpretability. Unlike a fully connected DNN (F-DNN), which computes the utility value of an alternative k by using the attributes of all the alternatives, ASU-DNN computes it by using only k's own attributes. Theoretically, ASU- DNN can substantially reduce the estimation error of F-DNN because of its lighter architecture and sparser connectivity, although the constraint of alternative-specific utility can cause ASU- DNN to exhibit a larger approximation error. Empirically, ASU-DNN has 2-3% higher prediction accuracy than F-DNN over the whole hyperparameter space in a private dataset collected in Singapore and a public dataset available in the R mlogit package. The alternative-specific connectivity is associated with the independence of irrelevant alternative (IIA) constraint, which as a domain-knowledge-based regularization method is more effective than the most popular generic-purpose explicit and implicit regularization methods and architectural hyperparameters. ASU-DNN provides a more regular substitution pattern of travel mode choices than F-DNN does, rendering ASU-DNN more interpretable.
    [Show full text]
  • Pseudo-Learning Effects in Reinforcement Learning Model-Based Analysis: a Problem Of
    Pseudo-learning effects in reinforcement learning model-based analysis: A problem of misspecification of initial preference Kentaro Katahira1,2*, Yu Bai2,3, Takashi Nakao4 Institutional affiliation: 1 Department of Psychology, Graduate School of Informatics, Nagoya University, Nagoya, Aichi, Japan 2 Department of Psychology, Graduate School of Environment, Nagoya University 3 Faculty of literature and law, Communication University of China 4 Department of Psychology, Graduate School of Education, Hiroshima University 1 Abstract In this study, we investigate a methodological problem of reinforcement-learning (RL) model- based analysis of choice behavior. We show that misspecification of the initial preference of subjects can significantly affect the parameter estimates, model selection, and conclusions of an analysis. This problem can be considered to be an extension of the methodological flaw in the free-choice paradigm (FCP), which has been controversial in studies of decision making. To illustrate the problem, we conducted simulations of a hypothetical reward-based choice experiment. The simulation shows that the RL model-based analysis reports an apparent preference change if hypothetical subjects prefer one option from the beginning, even when they do not change their preferences (i.e., via learning). We discuss possible solutions for this problem. Keywords: reinforcement learning, model-based analysis, statistical artifact, decision-making, preference 2 Introduction Reinforcement-learning (RL) model-based trial-by-trial analysis is an important tool for analyzing data from decision-making experiments that involve learning (Corrado and Doya, 2007; Daw, 2011; O’Doherty, Hampton, and Kim, 2007). One purpose of this type of analysis is to estimate latent variables (e.g., action values and reward prediction error) that underlie computational processes.
    [Show full text]
  • Lecture 18: Wrapping up Classification Mark Hasegawa-Johnson, 3/9/2019
    Lecture 18: Wrapping up classification Mark Hasegawa-Johnson, 3/9/2019. CC-BY 3.0: You are free to share and adapt these slides if you cite the original. Modified by Julia Hockenmaier Today’s class • Perceptron: binary and multiclass case • Getting a distribution over class labels: one-hot output and softmax • Differentiable perceptrons: binary and multiclass case • Cross-entropy loss Recap: Classification, linear classifiers 3 Classification as a supervised learning task • Classification tasks: Label data points x ∈ X from an n-dimensional vector space with discrete categories (classes) y ∈Y Binary classification: Two possible labels Y = {0,1} or Y = {-1,+1} Multiclass classification: k possible labels Y = {1, 2, …, k} • Classifier: a function X →Y f(x) = y • Linear classifiers f(x) = sgn(wx) [for binary classification] are parametrized by (n+1)-dimensional weight vectors • Supervised learning: Learn the parameters of the classifier (e.g. w) from a labeled data set Dtrain = {(x1, y1),…,(xD, yD)} Batch versus online training Batch learning: The learner sees the complete training data, and only changes its hypothesis when it has seen the entire training data set. Online training: The learner sees the training data one example at a time, and can change its hypothesis with every new example Compromise: Minibatch learning (commonly used in practice) The learner sees small sets of training examples at a time, and changes its hypothesis with every such minibatch of examples For minibatch and online example: randomize the order of examples for
    [Show full text]
  • Categorical Data
    Categorical Data Santiago Barreda LSA Summer Institute 2019 Normally-distributed data • We have been modelling data with normally-distributed residuals. • In other words: the data is normally- distributed around the mean predicted by the model. Predicting Position by Height • We will invert the questions we’ve been considering. • Can we predict position from player characteristics? Using an OLS Regression Using an OLS Regression Using an OLS Regression • Not bad but we can do better. In particular: • There are hard bounds on our outcome variable. • The boundaries affect possible errors: i.e., the 1 position cannot be underestimated but the 2 can. • There is nothing ‘between’ the outcomes. • There are specialized models for data with these sorts of characteristics. The Generalized Linear Model • We can break up a regression model into three components: • The systematic component. • The random component. • The link function. 푦 = 푎 + 훽 ∗ 푥 + 푒 The Systematic Component • This is our regression equation. • It specifies a deterministic relationship between the predictors and the predicted value. • In the absence of noise and with the correct model, we would expect perfect prediction. μ = 푎 + 훽 ∗ 푥 Predicted value. Not the observation. The Random Component • Unpredictable variation conditional on the fitted value. • This specifies the nature of that variation. • The fitted value is the mean parameter of a probability distribution. μ = 푎 + 훽 ∗ 푥 푦~푁표푟푚푎푙(휇, 휎2) 푦~퐵푒푟푛표푢푙푙(휇) Bernoulli Distribution • Unlike the normal, this distribution generates only values of 1 and 0. • It has only a single parameter, which must be between 0 and 1. • The parameter is the probability of observing an outcome of 1 (or 1-P of 0).
    [Show full text]
  • Mixed Pattern Recognition Methodology on Wafer Maps with Pre-Trained Convolutional Neural Networks
    Mixed Pattern Recognition Methodology on Wafer Maps with Pre-trained Convolutional Neural Networks Yunseon Byun and Jun-Geol Baek School of Industrial Management Engineering, Korea University, Seoul, South Korea {yun-seon, jungeol}@korea.ac.kr Keywords: Classification, Convolutional Neural Networks, Deep Learning, Smart Manufacturing. Abstract: In the semiconductor industry, the defect patterns on wafer bin map are related to yield degradation. Most companies control the manufacturing processes which occur to any critical defects by identifying the maps so that it is important to classify the patterns accurately. The engineers inspect the maps directly. However, it is difficult to check many wafers one by one because of the increasing demand for semiconductors. Although many studies on automatic classification have been conducted, it is still hard to classify when two or more patterns are mixed on the same map. In this study, we propose an automatic classifier that identifies whether it is a single pattern or a mixed pattern and shows what types are mixed. Convolutional neural networks are used for the classification model, and convolutional autoencoder is used for initializing the convolutional neural networks. After trained with single-type defect map data, the model is tested on single-type or mixed- type patterns. At this time, it is determined whether it is a mixed-type pattern by calculating the probability that the model assigns to each class and the threshold. The proposed method is experimented using wafer bin map data with eight defect patterns. The results show that single defect pattern maps and mixed-type defect pattern maps are identified accurately without prior knowledge.
    [Show full text]
  • Arxiv:1910.04465V2 [Cs.CV] 16 Oct 2019
    Searching for A Robust Neural Architecture in Four GPU Hours Xuanyi Dong1;2,Yi Yang1 1University of Technology Sydney 2Baidu Research [email protected], [email protected] Abstract 0 0 : input GDAS Conventional neural architecture search (NAS) ap- 3 : output on a DAG proaches are based on reinforcement learning or evolution- : sampled ary strategy, which take more than 3000 GPU hours to find : unsampled a good model on CIFAR-10. We propose an efficient NAS 1 approach learning to search by gradient descent. Our ap- proach represents the search space as a directed acyclic graph (DAG). This DAG contains billions of sub-graphs, each of which indicates a kind of neural architecture. To 2 avoid traversing all the possibilities of the sub-graphs, we develop a differentiable sampler over the DAG. This sam- pler is learnable and optimized by the validation loss af- ter training the sampled architecture. In this way, our ap- proach can be trained in an end-to-end fashion by gra- 3 dient descent, named Gradient-based search using Differ- entiable Architecture Sampler (GDAS). In experiments, we Figure 1. We utilize a DAG to represent the search space of a neu- can finish one searching procedure in four GPU hours on ral cell. Different operations (colored arrows) transform one node CIFAR-10, and the discovered model obtains a test error (square) to its intermediate features (little circles). Meanwhile, of 2.82% with only 2.5M parameters, which is on par with each node is the sum of the intermediate features transformed from the state-of-the-art.
    [Show full text]
  • Reinforcement Learning with Dynamic Boltzmann Softmax Updates Arxiv
    Reinforcement Learning with Dynamic Boltzmann Softmax Updates Ling Pan1, Qingpeng Cai1, Qi Meng2, Wei Chen2, Longbo Huang1, Tie-Yan Liu2 1IIIS, Tsinghua University 2Microsoft Research Asia Abstract Value function estimation is an important task in reinforcement learning, i.e., prediction. The Boltz- mann softmax operator is a natural value estimator and can provide several benefits. However, it does not satisfy the non-expansion property, and its direct use may fail to converge even in value iteration. In this paper, we propose to update the value function with dynamic Boltzmann softmax (DBS) operator, which has good convergence property in the setting of planning and learning. Experimental results on GridWorld show that the DBS operator enables better estimation of the value function, which rectifies the convergence issue of the softmax operator. Finally, we propose the DBS-DQN algorithm by applying dynamic Boltzmann softmax updates in deep Q-network, which outperforms DQN substantially in 40 out of 49 Atari games. 1 Introduction Reinforcement learning has achieved groundbreaking success for many decision making problems, including roboticsKober et al. (2013), game playingMnih et al. (2015); Silver et al. (2017), and many others. Without full information of transition dynamics and reward functions of the environment, the agent learns an optimal policy by interacting with the environment from experience. Value function estimation is an important task in reinforcement learning, i.e., prediction Sutton (1988); DEramo et al. (2016); Xu et al. (2018). In the prediction task, it requires the agent to have a good estimate of the value function in order to update towards the true value function.
    [Show full text]
  • Rethinking Feature Distribution for Loss Functions in Image Classification
    Rethinking Feature Distribution for Loss Functions in Image Classification Weitao Wan1∗ Yuanyi Zhong1,2∗† Tianpeng Li1 Jiansheng Chen1‡ 1Department of Electronic Engineering, Tsinghua University, Beijing, China 2Department of Computer Science, University of at Urbana-Champaign, Illinois, USA [email protected] [email protected] [email protected] [email protected] Abstract and the class label is used as the loss function. The soft- max loss has its probabilistic interpretation in that, for a We propose a large-margin Gaussian Mixture (L-GM) large class of distributions, the posterior distribution com- loss for deep neural networks in classification tasks. Dif- plies with the softmax transformation of linear functions of ferent from the softmax cross-entropy loss, our proposal the feature vectors [1]. It can also be derived from a binary is established on the assumption that the deep features of Markov Random Field or a Boltzmann Machine model [3]. the training set follow a Gaussian Mixture distribution. By However, the relationship between the affinity score and the involving a classification margin and a likelihood regular- probability distribution of the training feature space is vague. ization, the L-GM loss facilitates both a high classification In other words, for an extracted feature, its likelihood to the performance and an accurate modeling of the training fea- training feature distribution is not well formulated. ture distribution. As such, the L-GM loss is superior to the Several variants have been proposed to enhance the ef- softmax loss and its major variants in the sense that besides fectiveness of the softmax loss.
    [Show full text]