Determining a good feature set is critical to the performance of learning algorithms such as classifiers. Recently, researchers have proposed evolutionary-based feature extraction methods that aim to find a good feature set by combining the original features with new features generated by mathematical transformations of the original features. In this paper, we propose dynamically collecting past performance information on promising features and operators to use in our mutation method. We consider how to make our evolutionary algorithm more efficient and reliable by reducing overfitting. Preliminary results using UCI data show that our dynamic mutation method only slightly enhances the classification accuracy but it produces more reliable results.

Extracting good features is critical to the performance of learning algorithms such as classifiers. Feature extraction selects and transforms original features to find information hidden in data. Due to the huge search space of selection and transformation of features, exhaustive search is computationally prohibitive and randomized search such as evolutionary algorithms (EA) are often used. In our prior work on evolutionary-based feature extraction, an individual, which represents a set of features, is evaluated by estimating the accuracy of a classifier when the individual's feature set is used for learning. Although incorporating a learning algorithm during evaluation, which is called the wrapper approach, generally performs better than evaluating an individual simply by the statistical properties of data, which is called the filter appproach, our EA based on a wrapper approach suffers from overfitting, so that a slight enhancement of fitness in training can dramatically reduce the classification accuracy for unseen testing data. To cope with this problem, this paper proposes a two-population EA for feature extraction (TEAFE) that combines filter and wrapper approaches, and shows the promising preliminary results.

This paper proposes a method to recognize bare hand gestures using a dynamic vision sensor (DVS) camera. Different from conventional cameras, DVS cameras only respond to pixels with temporal luminance differences, which can greatly reduce the computational cost of comparing consecutive frames to track moving objects. Due to differences in available information, conventional vision techniques for gesture recognition may not be directly applicable in DVS based applications. This paper attempts to classify three different hand gestures made by a player during rock-paper-scissors game. We propose novel methods to detect the point where the player delivers a throw, to extract hand regions, and to extract useful features for machine learning based classification. Preliminary results show that our method produces enhanced accuracy of hand gesture recognition.

Differential evolution (DE) mutates solution vectors by the weighted difference of other vectors using arithmetic operations. As these operations cannot be directly extended to discrete combinatorial space, DE algorithms have been traditionally applied to optimization problems where the search space is continuous. In this paper, we use JADE, a self-adaptive DE algorithm, for winner determination in Combinatorial Auctions (CAs) where users place bids on combina tions of items. To adapt JADE to discrete optimization, we use a rank-based representation schema that produces only feasible solu tions and a regeneration operation that constricts the problem search space. It is shown that JADE compares favorably to a local stochastic search algorithm, Casanova, and a genetic algorithm based approach, SGA.

The rapid evolution of distributed sensor systems involving multiple heterogeneous sensors, the combination of mobile and stationary sensors, and dynamic information requirements presents a challenge for system resource utilization. On one hand, evolving standards and service oriented architecture concepts enables creation and use of multiple sensors and resources (including human reports). These technologies allow sensors and sources to be treated as Web services that can join the network at any time. On the other hand, traditional methods for optimizing these dynamic resources (e.g., classic optimization methods) are not feasible when the information requirements are rapidly changing, when there are multiple consumers of the requested information, and when the resources may dynamically appear (or disappear) from availability. This paper describes the concept of using a market-based approach for resource allocation. In this approach, sensors and network communications resources are treated as suppliers and human users (and data fusion algorithms) are treated as consumers. Dynamic auction methods are used to seek to match the dynamic information needs of the consumers with the capabilities of the suppliers. The "value" of the information to consumers is based on quantities such as accuracy, level of specificity, timeliness, etc. The paper describes the overall approach and presents results of numerical simulations.

The Web has become a worldwide repository of information which individuals, companies, and organizations utilize to solve or address various information problems. Many of these Web users utilize automated agents to gather this information for them. Some assume that this approach represents a more sophisticated method of searching. However, there is little research investigating how Web agents search for online information. In this research, we first provide a classification for information agent using stages of information gathering, gathering approaches, and agent architecture. We then examine an implementation of one of the resulting classifications in detail, investigating how agents search for information on Web search engines, including the session, query, term, duration and frequency of interactions. For this temporal study, we analyzed three data sets of queries and page views from agents interacting with the Excite and AltaVista search engines from 1997 to 2002, examining approximately 900,000 queries submitted by over 3,000 agents. Findings include: (1) agent sessions are extremely interactive, with sometimes hundreds of interactions per second (2) agent queries are comparable to human searchers, with little use of query operators, (3) Web agents are searching for a relatively limited variety of information, wherein only 18% of the terms used are unique, and (4) the duration of agent-Web search engine interaction typically spans several hours. We discuss the implications for Web information agents and search engines.

In this paper, we examine the relative forecast accuracy of information markets versus expert aggregation. We leverage a unique data source of almost 2000 people's subjective probability judgments on 2003 US National Football League games and compare with the "market probabilities" given by two different information markets on exactly the same events. We combine assessments of multiple experts via linear and logarithmic aggregation functions to form pooled predictions. Prices in information markets are used to derive market predictions. Our results show that, at the same time point ahead of the game, information markets provide as accurate predictions as pooled expert assessments. In screening pooled expert predictions, we find that arithmetic average is a robust and efficient pooling function; weighting expert assessments according to their past performance does not improve accuracy of pooled predictions; and logarithmic aggregation functions offer bolder predictions than linear aggregation functions. The results provide insights into the predictive performance of information markets, and the relative merits of selecting among various opinion pooling methods.