Abstract: Personality is defined as "the complex of all the attributes--behavioral, temperamental, emotional and mental--that characterize a unique individual" (Wordnet). Attempts to model those attributes in an artificial agent started in the 1960s with with personalities such as ELIZA, the psychotherapist and PARRY, the paranoid and have continued. This presentation will talk about the ideas behind artificial personalities and the attempt to incorporate personalities into the Spamalot anti-spam project.

Slides: Not available yet

Abstract: The emergence of new capabilities are observed in environments of human interaction. Specifically, a shared language emerges when diverse language groups interact for a consistently long period of time. This can be explained as a consequence of rationality (intelligence). Based on this, it is expected that agents that act rationally and have the capacity to learn would develop new capabilities when they interact with each other. Our research aims to study such emergent characteristics. Specifically, we intend to study emergent language characteristics when agents that communicate using different (formal) languages bargain with each other. 

This talk outlined some results from classical bargaining theory - which serves as a natural starting point for our research.

Slides: Not available yet

Abstract: Intelligent tutoring systems(ITSs) can provide benefits of one-on-one instruction automatically and cost effectively. To make the intelligent tutoring systems as effective as expert human tutors, the community is investigating methods to computationally model expert tutoring. Rules have been shown as an appropriate representation to model tutoring and can be easily applied to ITSs. In Recent years, researchers in ITS community start to use machine learning techniques to study tutoring and derive tutorial rules. This presentation introduces a method of learning tutorial rules from some annotated tutoring dialogues using CBA (Classification Based on Association). The tutorial rules include the ones for predicting tutor's move, the ones for predicting tutor's attitude and the ones for choosing tutoring knowledge content. Some preliminary experiment results are also presented. 

Note: Xin is a PhD student of Professor Barbara Di Eugenio, who is doing research in intelligent tutoring systems and natural language generation.

Slides: Not available yet

Abstract: Current spell checkers do a good job catching typos that result in garbled words, but fail to discover spelling mistakes that coincidentally end up being real words. Those errors cannot be found without taking context into account. To address this problem, a statistical language model is proposed, and some experimental results are presented. 

Note: Davide is a Ph.D. student of Prof Barbara Di Eugenio, doing research in natural language processing and intelligent tutoring systems.

Slides: Not available yet

Abstract: As the complexity of the computer system increases, the system becomes more prone to error and unmanageable. Finding a methodology to manage the ever-increasing complexity of computing environments has become a big challenge in the IT industry. The Autonomic Computing (AC) thus has been heavily studied in recent years to provide a solution for this problem. AC describes a computing model in which the system is self-healing, self-configured, self-protected and self-managed. Designed to mimic the human body's nervous system-in that the autonomic nervous system acts and reacts to stimuli independent of the individual's conscious input-an Autonomic Computing environment functions with a high level of artificial intelligence while remaining invisible to the users.

This presentation first gives an overview of the Autonomic computing (AC), focusing on the four fundamentals of AC, the AC architecture, and the Industry standards support; then it introduces the leading industry and university researches and projects on AC; at the end, the major research challenges of AC will be presented.

Slides: Available here in PDF format

Abstract: Randomness and random numbers are needed in a wide range of areas; they are of basic importance in computational statistics, in the implementation of probabilistic algorithms, and in related problems of statistical computing that have a stochastic ingredient like financial Modeling and artificial intelligence methods. In most applications, the actual relationship between successive points in a random sample has no physical significance; hence, randomness of the sample for approximating a uniform distribution is not critical.  Quasi random (low discrepancy) sequences are designed to have better uniformity and perform better than pseudo random sequences. For higher dimensional stochastic processes, however, the quality of the quasi random sequences rapidly decreases.  The focus of this research is to design optimal strategies for random number generations that will perform well in higher dimensions as well.

Since random numbers are used for various purposes it is important to define a performance index or indices for optimal random number generator. In this work, we are also proposing a novel and generalized approach for error characterization that is independent of probability distributions, number of variables, and/or functional form of the model. This approach is based on the fractal geometry.  In summary, this talk is a two part presentation where the first part will address the problem of optimal random number generator and the second part will be devoted to error characterization through fractal geometry.

Slides: Available here in PDF format

Abstract: Will a rational agent ever choose to deceive another agent?  This talk will present a situation in which a deliberative agent determines it is optimal to attempt to deceive another agent while playing the Multi-Agent Persistent Tiger Game.  Interactive Dynamic Influence Diagram will be used to demonstrate the planning process for deceptive agent.  Examining the agent's deliberative process will reveal a particular way in which rational agents will choose to adopt deceptive plans.

Slides: Not available yet

Abstract: The increasingly huge volume of personal information stored in a desktop computer is characterized by disparate models, unstructured contents, and implicit knowledge. Aiming at a semantic rich environment, a number of Semantic Desktop frameworks have been proposed, concentrating on different aspects, including organization, manipulation, and visualization of the data. In this project, we propose a layered and semantic ontology-based framework for personal information management, and we discuss its annotations, associations, and navigation. In particular, multiple ontologies are used to provide a uniform, conceptual model for the heterogeneous personal information, e.g., saved email messages, bibliography, photos, music, and video.  We also discuss query processing in two cases: query rewriting in a single personal information application, PIA, and that between two PIAs.

Slides: Not available yet

Abstract: Despite the widespread clinical use and acceptance of cochlear implants, the basic mechanisms that govern the response of the auditory nerve to electrical stimulation are not well understood.  The number of independent channels that can be used perceptually depends, in part, on the interactions between current fields around the electrodes. Computational models of current spread and of action-potential generation have been developed, but most of them ignore the current flow in the cochlea to spike generation in auditory nerve fibers. With the impedance data, this project created a 3 dimensional circuit model to simulate the current flow in the cochlear. A java application was developed based on this model to supply a 3D view of the simulation. The simulation results are consistent with experiment measurement.

Slides: Available here in PDF format

Abstract: In games such as Kriegspiel (a chess variant where players have no direct knowledge of the opponent's pieces' locations) the belief state’s sizes dwarf those of other partial information games -- like bridge, scrabble, and poker -- and there is no easy way to generate states satisfying the given observations. It has been shown that statistical sampling approaches can be developed to do well in such games. What we are trying to do is to introduce the notion of "adversary's model" in this framework. The idea is that, modeling the adversary via decision networks (influence diagrams), and keeping a belief over states and models, the statistical sampling method should work better than assuming random moves by the adversary.

Slides: Available here in PDF format

Abstract: Inter-vehicle communication by means of wireless Ad Hoc networking has the potential to improve traffic safety and comfort tremendously. Therefore, the application of Vehicular Ad Hoc Networks (VANETs) in the service of Intelligent Transportation Systems (ITS) has been highly focused in the recent years. Derived from the successful outcome of a cluster-based framework in Mobile Ad Hoc Networks (MANETs), we apply this network topology to VANETs. Unfortunately, previous studies lack realistic modeling of vehicle mobility and evaluation of clustering performance so they may not correlate well with performance in a real deployment. Hence, we propose a realistic micro-simulation model with the hope of contributing to clustering research in VANETs.  In this project, we analyze the effect of weighting two well-known clustering methods with the vehicle-specific position and velocity clustering logic to improve cluster stability over the simulation time.

Slides: Available here in PDF format

Abstract: Automatic discovering and predicting the hidden pattern or relationships among the monitoring data produced from the manufacturing and design processes are pivotal in improving the production quality. Among the existing data mining techniques, Evolutionary Computation methods are featured by taking an adaptive and parallel search procedure, conducting unconstrained approximation to the final solution (i.e., the mined knowledge from the data), and requiring minimum pre-knowledge about the problem domain. However, most traditional evolutionary computation procedures are only guided by the value fitness measurement of the evolved solutions, and the structure information about the solutions, which actually largely determines their functionality, is overlooked. This research aims at improving the Gene Expression Programming (GEP), a recently developed evolutionary computation algorithm, to fulfill complex data mining tasks by preserving and utilizing the self-emergence of solution structures during its evolutionary process. Our main hypotheses are (1) an efficient search procedure should be guided toward good solution structures; (2) decomposition of the functionality and hierarchical evolution of solutions are the ways to discover complex hidden knowledge. This talk will begin with the basic philosophy of Evolutionary Computation, followed by the introduction of our approaches, and ends up with some future work discussion.

Slides: Available here in PDF format

Abstract: With the expansion of computers and Internet into almost every corner of their workplaces and houses, people often find themselves overwhelmed by the tremendous amount of text documents (email, news articles, consumer product reviews, manuals, technical reports, journals, books, ...). Automatic text summarization is one tool which is used by people to deal with Information Overload. Previous studies of automatic text summarization rely on defining specific algorithms to produce summaries with respect to the characteristics of a specific domain. A summarizer designed in this way can produce good summaries in a restricted domain. However, the summarizer does not possess the ability to improve itself even in the domain it is specifically designed for or to adapt to another evaluation metric. In this research, a novel framework for machine summarization in which the summarizer learns from human-written summaries is proposed. In the first phase of our research, the summarizer aims to produce extractive summaries which on average have the best fitness to a certain evaluation metric. However, extractive summaries suffer from poor coherence, since the sentences in an extractive summary may be selected from different contexts and have no logical connections between them. Therefore, in the second phase of our study, we propose to solve this problem and produce near-abstractive summaries by using machine learning techniques in several steps, based on the output from the first phase. Specifically, we propose to use machine learning on sentence paraphrasing and condensation to produce summaries which are more coherent than the extractive ones. We present a road map to that goal and an execution plan for each step of our future study.

Slides: Available here in PDF format

Abstract: One of the most important tasks after a genome is sequenced and locations of genes within the genome are predicted, is to work out the possible functions of these genes. This talk will introduce the protein function prediction problem, and the need of using data mining techniques. New research progress on this topic and proposed work will be given in the talk as well.

Slides: Available here in PDF format

Abstract: Graphical models such as Bayesian networks, and Influence diagrams allow us to explicitly capture the structure of the domain i.e. causal relationships, conditional independences etc. Additionally, models such as dynamic Influence diagrams also represent an on-line method for computing the solution for frameworks like MDPs and POMDPs. However, traditionally, these models have concentrated on representing a world that is populated by a single agent only. In the past few years, several researchers have converged on the usefulness of these models in multiagent settings. Interesting examples of graphical models in multiagent settings are multiagent Influence diagrams (MAIDs), Influence diagram networks (IDNs), and Interactive Influence diagrams (IIDs). I will briefy survey each of these three models, with an emphasis on their solution techniques. Graphical modeling software such as Netica cannot represent or solve these models. However, one can, atleast in theory, use the Netica API to write a program that is capable of representing and solving these models. I will also talk about such implementation issues.

Slides: Multiagent Graphical Models.ppt

Abstract: Sony AIBO is an ideal platform of implementing AI algorithm (e.g. decision making). This presentation is going to explain the idea of realtime sensing about the environment and action performance in AIBO. It will show everything needed to implement an AIBO program.

Slides: AIAIBO.ppt

Abstract: Reasoning about others' beliefs and knowledge, i.e. the so called interactive beliefs and interactive knowledge respectively, has been the subject of continuous interest in several fields such as multiagent systems, artificial intelligence, game theory, and natural language applications. For e.g. in multiagent interaction, where an agent's optimal action is influenced by the actions of other agents present in the environment, the agent must reason about the beliefs of other agents about the environment, and what they believe about itself and so on. Consequently, hierarchies of beliefs arise in a natural and essential way in multiagent systems. Technically, these nestings may continue upto infinite levels making it difficult to reason with them, and therefore promoting approaches to truncate these belief nestings.

In addition to a discussion about such belief nestings, I will also present the concept of "common knowledge", which is one of the most well-studied example of infinite belief hierarchies. Common knowledge frequently arises in game theory and multiagent systems, and is therefore of interest to us.

Slides: nestedBelief.pdf