Transfer in variable-reward hierarchical reinforcement learning

Transfer learning seeks to leverage previously learned tasks to achieve faster learning in a new task. In this paper, we consider transfer learning in the context of related but distinct Reinforcement Learning (RL) problems. In particular, our RL problems are derived from Semi-Markov Decision Processes (SMDPs) that share the same transition dynamics but have different reward functions that are linear in a set of reward features. We formally define the transfer learning problem in the context of RL as learning an efficient algorithm to solve any SMDP drawn from a fixed distribution after experiencing a finite number of them. Furthermore, we introduce an online algorithm to solve this problem, Variable-Reward Reinforcement Learning (VRRL), that compactly stores the optimal value functions for several SMDPs, and uses them to optimally initialize the value function for a new SMDP. We generalize our method to a hierarchical RL setting where the different SMDPs share the same task hierarchy. Our experimental results in a simplified real-time strategy domain show that significant transfer learning occurs in both flat and hierarchical settings. Transfer is especially effective in the hierarchical setting where the overall value functions are decomposed into subtask value functions which are more widely amenable to transfer across different SMDPs.     

Inductive transfer with context-sensitive neural networks

Context-sensitive Multiple Task Learning, or csMTL, is presented as a method of inductive transfer which uses a single output neural network and additional contextual inputs for learning multiple tasks. Motivated by problems with the application of MTL networks to machine lifelong learning systems, csMTL encoding of multiple task examples was developed and found to improve predictive performance. As evidence, the csMTL method is tested on seven task domains and shown to produce hypotheses for primary tasks that are often better than standard MTL hypotheses when learning in the presence of related and unrelated tasks. We argue that the reason for this performance improvement is a reduction in the number of effective free parameters in the csMTL network brought about by the shared output node and weight update constraints due to the context inputs. An examination of IDT and SVM models developed from csMTL encoded data provides initial evidence that this improvement is not shared across all machine learning models.     

Learning to Locate Informative Features for Visual Identification

Object identification is a specialized type of recognition in which the category (e.g. cars) is known and the goal is to recognize an object’s exact identity (e.g. Bob’s BMW). Two special challenges characterize object identification. First, inter-object variation is often small (many cars look alike) and may be dwarfed by illumination or pose changes. Second, there may be many different instances of the category but few or just one positive “training” examples per object instance. Because variation among object instances may be small, a solution must locate possibly subtle object-specific salient features, like a door handle, while avoiding distracting ones such as specular highlights. With just one training example per object instance, however, standard modeling and feature selection techniques cannot be used. We describe an on-line algorithm that takes one image from a known category and builds an efficient “same” versus “different” classification cascade by predicting the most discriminative features for that object instance. Our method not only estimates the saliency and scoring function for each candidate feature, but also models the dependency between features, building an ordered sequence of discriminative features specific to the given image. Learned stopping thresholds make the identifier very efficient. To make this possible, category-specific characteristics are learned automatically in an off-line training procedure from labeled image pairs of the category. Our method, using the same algorithm for both cars and faces, outperforms a wide variety of other methods.     

Unsupervised Learning of Human Action Categories Using Spatial-Temporal Words

We present a novel unsupervised learning method for human action categories. A video sequence is represented as a collection of spatial-temporal words by extracting space-time interest points. The algorithm automatically learns the probability distributions of the spatial-temporal words and the intermediate topics corresponding to human action categories. This is achieved by using latent topic models such as the probabilistic Latent Semantic Analysis (pLSA) model and Latent Dirichlet Allocation (LDA). Our approach can handle noisy feature points arisen from dynamic background and moving cameras due to the application of the probabilistic models. Given a novel video sequence, the algorithm can categorize and localize the human action(s) contained in the video. We test our algorithm on three challenging datasets: the KTH human motion dataset, the Weizmann human action dataset, and a recent dataset of figure skating actions. Our results reflect the promise of such a simple approach. In addition, our algorithm can recognize and localize multiple actions in long and complex video sequences containing multiple motions.     

Efficient Learning of Relational Object Class Models

We present an efficient method for learning part-based object class models from unsegmented images represented as sets of salient features. A model includes parts’ appearance, as well as location and scale relations between parts. The object class is generatively modeled using a simple Bayesian network with a central hidden node containing location and scale information, and nodes describing object parts. The model’s parameters, however, are optimized to reduce a loss function of the training error, as in discriminative methods. We show how boosting techniques can be extended to optimize the relational model proposed, with complexity linear in the number of parts and the number of features per image. This efficiency allows our method to learn relational models with many parts and features. The method has an advantage over purely generative and purely discriminative approaches for learning from sets of salient features, since generative method often use a small number of parts and features, while discriminative methods tend to ignore geometrical relations between parts. Experimental results are described, using some bench-mark data sets and three sets of newly collected data, showing the relative merits of our method in recognition and localization tasks.     

Supervised and Reinforcement Evolutionary Learning for Wavelet-based Neuro-fuzzy Networks

This study presents a wavelet-based neuro-fuzzy network (WNFN). The proposed WNFN model combines the traditional Takagi–Sugeno–Kang (TSK) fuzzy model and the wavelet neural networks (WNN). This study adopts the non-orthogonal and compactly supported functions as wavelet neural network bases. A novel supervised evolutionary learning, called WNFN-S, is proposed to tune the adjustable parameters of the WNFN model. The proposed WNFN-S learning scheme is based on dynamic symbiotic evolution (DSE). The proposed DSE uses the sequential-search-based dynamic evolutionary (SSDE) method. In some real-world applications, exact training data may be expensive or even impossible to obtain. To solve this problem, the reinforcement evolutionary learning, called WNFN-R, is proposed. Computer simulations have been conducted to illustrate the performance and applicability of the proposed WNFN-S and WNFN-R learning algorithms.     

Asynchronous action-reward learning for nonstationary serial supply chain inventory control

Action-reward learning is a reinforcement learning method. In this machine learning approach, an agent interacts with non-deterministic control domain. The agent selects actions at decision epochs and the control domain gives rise to rewards with which the performance measures of the actions are updated. The objective of the agent is to select the future best actions based on the updated performance measures. In this paper, we develop an asynchronous action-reward learning model which updates the performance measures of actions faster than conventional action-reward learning. This learning model is suitable to apply to nonstationary control domain where the rewards for actions vary over time. Based on the asynchronous action-reward learning, two situation reactive inventory control models (centralized and decentralized models) are proposed for a two-stage serial supply chain with nonstationary customer demand. A simulation based experiment was performed to evaluate the performance of the proposed two models. Chang Ouk Kim received his Ph.D. in industrial engineering from Purdue University in 1996 and his B.S. and M.S. degrees from Korea University, Republic of Korea in 1988 and 1990, respectively. From 1998--2001, he was an assistant professor in the Department of Industrial Systems Engineering at Myongji University, Republic of Korea. In 2002, he joined the Department of Information and Industrial Engineering at Yonsei University, Republic of Korea and is now an associate professor. He has published more than 30 articles at international journals. He is currently working on applications of artificial intelligence and adaptive control theory in supply chain management, RFID based logistics information system design, and advanced process control in semiconductor manufacturing. Ick-Hyun Kwon is a postdoctoral researcher in the Department of Civil and Environmental Engineering at University of Illinois at Urbana-Champaign. Previous to this position, Dr. Kwon was a research assistant professor in the Research Institute for Information and Communication Technology at Korea University, Seoul, Republic of Korea. He received his B.S., M.S., and Ph.D. degrees in Industrial Engineering from Korea University, in 1998, 2000, and 2006, respectively. His current research interests are supply chain management, inventory control, production planning and scheduling. Jun-Geol Baek is an assistant professor in the Department of Business Administration at Kwangwoon University, Seoul, Korea. He received his B.S., M.S., and Ph.D. degrees in Industrial Engineering from Korea University, Seoul, Korea, in 1993, 1995, and 2001 respectively. From March 2002 to February 2007, he was an assistant professor in the Department of Industrial Systems Engineering at Induk Institute of Technology, Seoul, Korea. His research interests include machine learning, data mining, intelligent machine diagnosis, and ubiquitous logistics information systems. An erratum to this article can be found at     

Verification of evolving software via component substitutability analysis

This paper presents an automated and compositional procedure to solve the substitutability problem in the context of evolving software systems. Our solution contributes two techniques for checking correctness of software upgrades: (1) a technique based on simultaneous use of over-and under-approximations obtained via existential and universal abstractions; (2) a dynamic assume-guarantee reasoning algorithm—previously generated component assumptions are reused and altered on-the-fly to prove or disprove the global safety properties on the updated system. When upgrades are found to be non-substitutable, our solution generates constructive feedback to developers showing how to improve the components. The substitutability approach has been implemented and validated in the ComFoRT reasoning framework, and we report encouraging results on an industrial benchmark. This is an extended version of a paper, Dynamic Component Substitutability Analysis, published in the Proceedings of the Formal Methods 2005 Conference, Lecture Notes in Computer Science, vol. 3582, by the same authors. This research was sponsored by the National Science Foundation under grant nos. CNS-0411152, CCF-0429120, CCR-0121547, and CCR-0098072, the Semiconductor Research Corporation under grant no. TJ-1366, the US Army Research Office under grant no. DAAD19-01-1-0485, the Office of Naval Research under grant no. N00014-01-1-0796, the ICAST project and the Predictable Assembly from Certifiable Components (PACC) initiative at the Software Engineering Institute, Carnegie Mellon University. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of any sponsoring institution, the US government or any other entity.     

The Impact of Interaction and Social Learning on Aggregate Expectations

We examine the effect of information sharing within small world networks. Agents receive a signal correlated with the state of the world (SoW) which is adjusted following discussions with neighbours. If one agent in the network, referred to as an expert, does not engage in social learning (that is they always follow their own signal) then all agents learn the SoW. It is found that volatility in the mean level of expectations varies with changes in the number of experts and the network structure. A trade-off emerges between the level of volatility and the speed at which agents learn of changes to the SoW. A second finding is that certain network structures lead to information cascades.      

A neural network job-shop scheduler

This paper focuses on the development of a neural network (NN) scheduler for scheduling job-shops. In this hybrid intelligent system, genetic algorithms (GA) are used to generate optimal schedules to a known benchmark problem. In each optimal solution, every individually scheduled operation of a job is treated as a decision which contains knowledge. Each decision is modeled as a function of a set of job characteristics (e.g., processing time), which are divided into classes using domain knowledge from common dispatching rules (e.g., shortest processing time). A NN is used to capture the predictive knowledge regarding the assignment of operation’s position in a sequence. The trained NN could successfully replicate the performance of the GA on the benchmark problem. The developed NN scheduler was then tested against the GA, Attribute-Oriented Induction data mining methodology and common dispatching rules on a test set of randomly generated problems. The better performance of the NN scheduler on the test problem set compared to other methods proves the feasibility of NN-based scheduling. The scalability of the NN scheduler on larger problem sizes was also found to be satisfactory in replicating the performance of the GA.