Artificial Intelligence research at IBM today is at the forefront of many of the hottest areas. We are pursuing research in a wide variety of methodologies, including machine learning, Bayesian reasoning, knowledge representation, logic programming, common sense reasoning and planning, etc. We are particularly interested in synthesizing technologies, whether combining AI methodologies or integrating AI techniques with other technologies. Our focus is on using AI to solve challenging technical and commercial problems, and to advance the state-of-the-art in many areas, such as electronic commerce, supply chain management, autonomic computing, and exploratory vision. Research is being conducted either in core AI technologies, or its applications, in most of IBM Research labs, including Watson, Almaden, Haifa, India, Beijing, Tokyo, and Zurich.

Since many web-based business transactions are automated through intelligent agents, it is necessary to understand information economies, large systems of software agents that perform Internet transactions. We study the role of intelligent agents in setting prices in situations such as auctions and negotiations. Such situations can vary in the amount of knowledge that can be assumed and the amount of reasoning that the participants will perform. We have studied different strategies based on different assumptions and have formally analyzed the behavior of corresponding algorithms. We have implemented PriceBots and ShopBots, intelligent agents that can be programmed to implement particular strategies for retailers or consumers.

We have introduced techniques for difficult planning and scheduling problems. For example, we have examined the problem of scheduling paper manufacturing jobs when such factors as sequence-dependent setups, job-machine restrictions, batch size preferences, and downstream scheduling consequences must be considered simultaneously. Optimal solutions, even when only considering minimizing total tardiness on one machine, can be NP-hard. Using the A-Team architecture, in which planning agents cooperate by exchanging results, we have successfully implemented algorithms that our experiments show gives impressive and computationally efficient results.

We are also developing a framework that allows the building of hybrid agents that can perform a variety of intelligent activities. ABLE (Agent Building and Learning Environment) focuses on building hybrid agents that can both reason and learn. The ABLE framework consists of a set of core JavaBeans™ and a set of function-specific JavaBeans. We intend to make this platform compatible with the FIPA (Foundation for Intelligent Physical Agents) standards in Java.

In order to guarantee effectiveness, the performance of such services as web transactions and electronic mail needs to be managed. We seek to increase the level of automation for performance management. Our work includes the study of predictive detection, which gives advanced knowledge of performance degradations, and event mining, a particular form of data mining that recognizes situations that require action to ensure good performance. The Real-time Active Inference and Learning (RAIL) project aims to develop efficient techniques for real-time inference (diagnosis and prognosis) and learning (model adaptation to system changes) in complex distributed systems. An important property that differentiates this approach from ''passive'' data analysis is its ability to perform an active, online selection and execution of tests and measurements for more cost-efficient reasoning and learning.

We are examining the ways in which a knowledge of basic common-sense facts and an ability to reason with those facts can enhance interaction with automated systems such as online banking services. For example, it is obvious to people — but not necessarily to computers — that a person who is buying a house will probably need a mortgage and insurance, and possibly a car. We are using semantic networks to allow this sort of reasoning. In another research effort, we are developing MeanX, which focuses on modeling cognitive processes such as representing knowledge, reasoning, and learning on the symbolic level, which has led to the development of a system that understands meaning in natural language.

In order for a computer to understand the image that a camera produces, it must have a representation that it can manipulate, along with the background knowledge, context, and computational and inferential methods that it needs to understand the image. We are learning how vision systems can be used in laboratories, to track the motion of laboratory animals; in supermarkets, to recognize and classify produce; and in biometric applications, to recognize and classify faces and fingerprints.

We are also involved in researching various aspects of knowledge representation in many of our labs. In one project, we have successfully conducted an experiment in integrating CYC into an advanced natural language question answering system, while in another, we are exploring the integration of information extraction with hybrid reasoners, in collaboration with Stanford's Knowledge Systems Lab. we have also taken a lead role in the development of several standards, such as OWL, and have developed a methodology for evaluating an ontology using formal evaluation criteria.