Session I (5602 Beckman)

Speaker: Clay Holroyd, Neuroscience Program
Time: 1:00pm
Title: Is the Error-Related Negativity Generated by a Dopaminergic Error Signal for Reinforcement Learning? Hypothesis, Model, and Experiment.
Author(s): Clay Holroyd, Jesse Reichler, and Michael G. H. Coles
Abstract: The Error-related Negativity (ERN) is a fronto-centrally distributed component of the event-related brain potential (ERP) elicited when human subjects make errors in a variety of experimental tasks. The ERN is associated with the activity of a flexible error processing system and appears to be generated in the anterior cingulate cortex. Computational models suggest that the mesencephalic dopamine system may mediate a particular kind of reinforcement learning signal called a "temporal difference error." In this talk, we hypothesize that the ERN is elicited by an error signal carried to the anterior cingulate by the mesencephalic dopamine system, and explore the ways in which the ERN does and does not behave like a TD error. We propose that while a subject engages in an experimental task, the basal ganglia continually monitor motor activity, external cues and feedback, and make ongoing predictions about whether each trial will end in success or failure. When a positive or neutral estimate of a trial's outcome is revised in favor of a prediction of failure, a negative error signal is referred via the dopaminergic projection to frontal motor areas, where an ERN is generated. This hypothesis is formalized in a computational model based on the method of temporal differences. We compare and contrast several predictions made by the model with data from an event-related potential experiment.

Speaker: Mark Brodie, Department of Computer Science
Time: 1:10pm
Title: An Introduction to Iterated Phantom Induction
Author(s): Mark Brodie and Gerald DeJong
Abstract: We advance a machine learning method called iterated phantom induction which incorporates domain knowledge into the learning process in a flexible and general way. Domain knowledge which is expressed naturally by a human expert can be utilized by the learning system without needing to change the underlying machine learning algorithm. This method yields efficient and robust learning in a variety of domains. In particular, the method is applied to a challenging bicycle riding task and is compared with the performance of reinforcement learning.
Slides: for this talk (PostScript)
Paper(s): related papers

Speaker: Harlan Harris, Department of Computer Science
Time: 1:20pm
Title: An Initial-Weights Bias in Sequence-Producing Elman Networks
Author(s): Harlan Harris
Abstract: This work describes how a well-known and generally beneficial bias in feedforward networks, the restricted distribution of the initial weights, can cause a learning failure in an sequence-producing Elman recurrent network. The types and tasks of recurrent networks are reviewed, then a very simple task is explored. The analysis of the performance of a set of networks trained on this task with varying initial conditions shows that the minima in weight space change drastically during training, and that typically-used small initial weights prevent this class of network from escaping a catastrophic local minima.

Speaker: Ole Jakob Mengshoel, Department of Computer Science
Time: 1:30pm
Title: Probabilistic Crowding: Deterministic Crowding with Probabilistic Replacement
Author(s): Ole Jakob Mengshoel, David E. Goldberg
Abstract: This talk presents a novel niching algorithm, probabilistic crowding. Like its predecessor, deterministic crowding, probabilistic crowding is fast, simple, and requires no parameters beyond that of the classical GA. In probabilistic crowding, subpopulations are maintained reliably, and it is possible to analyze and predict how this maintenance takes place.

Speaker: William H. Hsu, National Center for Supercomputing Applications (NCSA)
Time: 1:40pm
Title: High-Performance Neural, Bayesian, and Genetic Computation for Scalable Data Mining: Research at NCSA
Author(s): William H. Hsu, Loretta Auvil, William M. Pottenger, Tom Redman, David Tcheng, Michael Welge, Jie Wu, and Ting-Hao Yang
Abstract: This talk addresses the problem of developing scalable techniques for learning from data. Improved data gathering instrumentation and practices have lead to a rapid increase in the volume of data available for analytical applications, which has not been matched by improved computational methods for discovering statistical models from data and applying them inferentially.
    I will give a brief synopsis of research on the use of machine learning methods to "scale up" data mining and knowledge discovery methods to large, real-world applications. The emphasis is on improving the accuracy and efficiency of learning systems (such as artificial neural networks and Bayesian networks) by partitioning, selecting, and synthesizing data channels (attributes) to formulate more manageable subproblems. The talk will conclude with a short, high-level overview of D2K, an implemented system for rapid knowledge discovery in databases (KDD).
Slides: full-length version of this talk (Microsoft PowerPoint 97)
Paper(s): related papers
 

Session II-A: General Track (5602 Beckman)

Speaker: Les Gasser, Graduate School of Library and Information Sciences (GSLIS)
Time: 2:05pm
Title: UIUC Agents and Multi-agents Systems Group (AMAG)
Author(s): Professors Gul Agha (CS), Les Gasser (GSLIS), Mike Shaw (Commerce/Business); John Barron, Nadeem Jamali, Eric Rankin, Christoph Schlueter, Gek-Woo Tan, Prasanna Thati, Carlos Varela, James Waldby
Abstract: The UIUC Agents and Multi-Agents Group (AMAG) is investigating theories, methods, technologies, and applications of agent-oriented and multi-agent information systems. Generally, "agents" are computational objects with characteristics like autonomy; structured, persistent action; sophisticated, resource-bounded reasoning; integrated architectures; and adaptive, learning capabilities, that act in complex, dynamic environments. "Social agents" interact with others (and with people), and may have aspects of emotion; personality; believability; multimodal interaction, and so on. "Multi-agent systems" (MAS) are collective, agent-based systems, possibly involving people as participants. Hard MAS problems include dynamic division of labor; coordination of interdepentent activities; structures for teams, groups, and organizations; and managing conflict and heterogeneity. Engineering, deploying, and evaluating agent-based and multi-agent systems are active areas of research. Practical application areas for MAS include electronic commerce; information gathering and management; distributed resource allocation; simulations of social and biological systems; muti-agent robotics; interacting personal digital assistants and information appliances, etc. UIUC-AMAG members are investigating programming and resource allocation models for distributed and mobile agents; market based resource allocation; security issues for agents and mobile agents; the nature of social knowledge and organizational memory; language evolution in agent communities; agent-based networked information systems and information management; multi-agent enterprise information modeling and supply-chain analysis; genetic algorithm models of group decision and learning; business-to-business electronic commerce; and coordination models for concurrent activities.

Speaker: Martin Pelikan, Department of General Engineering
Time: 2:15pm
Title: The Bayesian Optimization Algorithm
Author(s): Martin Pelikan, David E. Goldberg
Abstract: Recently, in evolutionary computation there has been a growing interest in linkage learning, or the identification of building blocks in a problem by the algorithm on the fly and the use of the acquired information to guarantee a proper growth and mixing of building blocks. However, most of the proposed methods extending the genetic algorithms in order to solve the difficulties of genetic algorithms to mix and reproduce the building blocks efficiently without any prior information of the problem were either insufficiently powerful, efficient, or they required some problem specific knowledge in order to be applicable.
    The so-called Bayesian optimization algorithm (BOA), uses techniques for learning Bayesian networks in order to model the promising solutions. Instead of applying crossover to randomly picked pairs of promising solutions, the joint distribution encoded by the constructed network is used to generate new candidate solutions. The algorithms proved to be very efficient for decomposable problems with or without overlapping. It is able to solve most of the tested problems in a close to linear time.
    In this talk, I will present the basic principles of the algorithms based on probabilistic modeling of promising solutions and briefly describe the basic principle of the BOA.

Speaker: Paul Patton, Neuroscience Program
Time: 2:25pm
Title: A Probabilistic Interpretation of Multisensory Enhancement in the Superior Colliculus
Author(s): Thomas J. Anastasio, Paul E. Patton, and Kamel Belkacem-Boussaid
Abstract: The deep layers of the superior colliculus (SC) integrate multimodal sensory inputs and initiate an orienting response to the stimulus source (target). Enhancement is a form of multisensory integration in which the neural response to a stimulus of one modality is augmented by a stimulus of another modality. Enhancement exhibits inverse effectiveness: combinations of weaker unimodal stimuli produce more enhancement. These findings can be accounted for by the hypothesis that the response of a deep SC neuron is proportional to the probability that a target is present given its sensory inputs. Sensory inputs are treated as stochastic, with spikes being more probable when a target is present. The mean difference between target-present and target-absent activity is small for ambiguous and large for unambiguous inputs. Given such inputs, the probability that a target is present can be computed using Bayes’s rule. This model provides interpretations for both multimodal enhancement and inverse effectiveness. The model suggests that the function of multisensory enhancement is to disambiguate ambiguous unimodal inputs. It suggests that inverse effectiveness results because the increase in target probability due to multimodal integration is greater when the unimodal inputs are weaker. (Supported by NSF Grant (IBN 92-21823) and the Illinois Critical Research Initiative.)

Speaker: Kamel Belkacem-Boussaid, Neuroscience Program
Time: 2:35pm
Title: A New Image Smoothing Method Based on a Simple Model of Spatial Processing in the Early Stages of Human Vision
Author(s): Kamel Belkacem-Boussaid
Abstract: The difficulty of preserving edges is central to the problem of smoothing images. The main problem is that of distinguishing between meaningful contours and noise, so that the image can be smoothed without loss of details. Substantial efforts have been devoted to solving this difficult problem, and a plethora of filtering methods have been proposed in the literature. Nonlinear filters have proved to be more efficient than their linear counterparts. Here, a new nonlinear filter for noise smoothing is introduced. This filter is based on the psychophysical phenomenon of human visual contrast sensitivity. Results on real images are presented to demonstrate the validity of our approach compared to other known filtering methods.

Speaker: Malcolm MacIver, Neuroscience Program
Time: 2:39pm
Title: Design of an Electro-orienting Fish Robot
Author(s): Malcolm MacIver
Abstract: Electrosense allows some unique fish to hunt, navigate, and communicate in complete darkness using a self-generated electric field. These fish generate a weak, sinusoidal electric discharge of around 1 kHz. Objects that differ in impedance from the surrounding water cause amplitude modulations of this carrier which are transduced by electroreceptors in the skin, modulating the high baseline firing rate of the primary afferents.
    We will be combining detailed computational models of the various components of the sensory pathway and the electrosensory milieu into a hardware implementation of an electro-orienting response. Placement of an object that differs in impedance from the surrounding water near the "robotic" fish would result in rotation of the fish toward the stimulus until the distance from the nearest sensor to the object is minimized. I will briefly discuss some design issues in going from biology to neuromorphic implementation.
 

Session II-B: Applications Track (4269 Beckman)

Speaker: Vivek Pinto, Department of Business Administration and NCSA
Time: 2:05pm
Title: Finite-Capacity, Multi-Objective Optimization by Simulated Annealing
Author(s): Vivek Pinto, William M. Pottenger, and William "Tilt" Thompkins
Abstract: The advent of information technology has given rise to enterprise wide information systems that enhance the flow of critical information through all levels of an enterprise. The chaotic nature of the environment that they operate has driven the search for models that can simulate the firm’s behavior over a competitive state space. Competitive pressures have forced business to seek competitive advantages and optimal cost efficiencies. The presence of multiple objectives with interdependent stochastic determinants makes the search for these optima extremely complex. This paper describes an attempt to build an enterprise wide model for the US Air Force C17 transport fleet. The salient features include multi-objective, multi-level optimization using simulated annealing on different temporal cycles, with finite capacities and predetermined policy constraints such as the funding and readiness levels. It is also an attempt to apply Monte Carlo simulation techniques to such problems.
Slides: for this talk (Microsoft PowerPoint 97)

Speaker: Barbara Minsker, Department of Civil and Environmental Engineering
Time: 2:10pm
Title: Challenges in Water Quality Management
Author(s): Barbara Minsker et al
Abstract: My research group has been investigating applications of genetic algorithms (GAs) to groundwater management problems.
    The GAs are usually coupled with numerical simulation models, which determine the fitness of candidate management designs. For complex, field-scale sites, the computational effort of evaluating each design is prohibitive. Current research is focusing on improving computational effort through three main approaches: (1) hierarchical solution of the problem, with a lower tier of GA population(s) generating "first-cut" solutions using simpler analytical models or coarse-mesh numerical approximations, which are then polished by higher tiers; (2) hybrid genetic algorithms, which couple Gas with gradient search routines; and (3) developing efficient sampling schemes for noisy GAs, which consider uncertainty in model parameters.
    For the future, I would like to investigate how evolutionary computing and/or artificial neural networks can be used to enhance decision making. Presently, millions of dollars are being spent to collect substantial quantities of data at contaminated sites worldwide. However, only a small fraction of this information is actually used in the decision making process, primarily because existing models can only consider a limited number of quantitative parameters. Substantial quantitative and qualitative information also exists that should be considered. I would be interested in brainstorming some ideas with those more familiar with the ANN/EC theory, to see how we could take advantage of these data to improve decision making. The simplest approach would be to simply throw all of the information into something like an artificial neural network, developing a new predictive tool that would replace current simulation models. However, simulation models consider important relationships such as conservation of mass, momentum, and energy, and any predictive tool that completely ignores those relationships seems doomed to failure, or at least little acceptance in the field. I would like to explore how simulation models could be used with ANN/EC to develop enhanced predictive tools that can be used for better decision making.

Speaker: Scott Burns, Department of General Engineering
Time: 2:15pm
Title: GA Application to Construction Planning and Structural Design--Synopsis of Recent and Current Research
Author(s): Scott Burns, Liang Liu, Chung-Wei Feng; Scott Burns, Keith Mueller
Abstract:
    Construction Planning – GA has been applied to finding the optimal construction time-cost trade-off curve for a project with multiple activities and multiple options for performing each activity. The fitness function is related to the distance from the convex hull that is formed by the population.
    Structural Design – It is possible to proportion a frame structure in many different ways to achieve an efficient fully-stressed design. Typical frame structures with multiple loading conditions may have dozens of potential fully-stressed designs. Many of these designs act as repellers under the action of standard design processes, and are therefore difficult to locate. A niching GA approach is proposed to locate all fully-stressed designs for a given frame structure.

Speaker: David Dubin, Graduate School of Library and Information Sciences (GSLIS)
Time: 2:20pm
Title: Evaluation Strategies for Large-Scale Neural Network Applications
Author(s): David Dubin
Abstract: Kohonen’s Self-Organizing Map is a very popular connectionist model, but few published studies describe the application of well-known cluster validation methods to the SOM. A few studies have compared the SOM to other classification methods, but usually with assumptions that aren’t consistent with the ways SOMs are used in large-scale applications. Furthermore, SOM proponents may have the mistaken impression that problems of scale rule out traditional validation methods.
    In scaling validation techniques for large SOM applications, the nature of the clusters and how they are apprehended by the SOM are at least as important as algorithm complexity and the availability of computing resources. We describe extensions to existing algorithms for generating artificial test clusters, aimed at supporting validation studies for large-scale SOMs. We propose ways in which other validation techniques (such as those based on the random graph hypothesis) may be employed in large-scale SOM applications.

Speaker: Fernando Lobo, Department of General Engineering
Time: 2:25pm
Title: Genetic Algorithms for Electrical Network Expansion Planning: A Demonstration
Author(s): Fernando Lobo
Abstract: This talk illustrates the application of a genetic algorithm (GA) to the problem of electrical distribution network expansion planning, along with a user-friendly GA application to solve it. The resulting application is flexible, easy to use, and requires no knowledge of GAs from the user. During the GA simulation, the best solution encountered so far is displayed in an animation-like fashion. At the end, the user is allowed to change the final solution by using his own expertise about the problem. The combination is a hybrid system. The GA finds a good region of the search space, and the decision-maker uses domain knowledge to tune the solution found by the GA. The main purpose of this work was to convince managers and engineers at the Portuguese National Electrical Utility ("Electricidade de Portugal – EDP") that genetic algorithms work, and have potential to solve difficult optimization problems of interest to the company. In addition, this work has an educational flavor and can be used in the classroom as a teaching tool to give students a first experience with a genetic algorithm.

Speaker: Anne Raich, Department of Civil Engineering
Time: 2:30pm
Title: Implicit Redundant Representation Genetic Algorithms for Synthesis in Conceptual Design
Author(s): Anne Raich
Abstract: Performing synthesis during conceptual design provides substantial cost savings by selecting the structural topology and geometry of the design, in addition to selecting the member sizes. Traditional optimization methods, however, cannot synthesis design solutions that have diverse structural topologies and geometries. A new evolutionary-based search method will be presented that supports the synthesis of design solutions in unstructured problem domains. The implicit redundant representation genetic algorithm (IRR GA) uses redundancy to support the self-organization of the representation by encoding a variable number of location independent design parameters. Using an unstructured definition of the problem domain allows the definition and evaluation of diverse structures that had variable topologies and geometries. The IRR GA provides several benefits: redundant segments protected existing variables from the disruption of crossover and mutation; new variables could be designated within previously redundant segments; and the number of parameters represented dynamically changed the dimensions of the search space. The IRR GA does not require the definition of a ground structure or heuristic rules to add or remove structural elements. Experimental results obtained by applying the IRR GA to evolve synthesis design solutions for an unstructured, multi-objective frame problem domain will be presented. Several levels of unstructured formulations have been examined. The difficulties of evaluating the topology and geometry dependent loading configurations, penalty functions, and multiple objectives will be discussed. Several of the novel frame designs generated by the IRR GA synthesis design method will be shown that compare favorably with solutions obtained using a trial and error design process.

Speaker: Lei Tian, Department of Agricultural Engineering
Time: 2:35pm
Title: An Environmentally Adaptive Segmentation Algorithm (EASA) for Aerial Images
Author(s): Lei Tian
Abstract: An environmentally adaptive segmentation algorithm (EASA) was developed for outdoor field plant detection. Based on a partially supervised learning process, the algorithm can learn from environmental conditions in outdoor agricultural fields and build an image segmentation look-up table on-the-fly. Experiments showed that the algorithm can adapt to most daytime conditions in outdoor fields, such as changes in light source temperature and soil type. When compared to a static segmentation technique which was trained under sunny conditions, the EASA improved the image segmentation by correctly classifying 26.9% and 54.3% more object pixels under partially cloudy and overcast conditions, respectively.

Speaker: Halil Ceylan, Department of Civil Engineering
Time: 2:39pm
Title: An Artificial Neural Network Model for Finite-Element Stress Analysis in Aircraft Pavement Design
Author(s): Halil Ceylan
Abstract: An artificial neural network (ANN) model has been trained with the results of ILLI-SLAB finite element program and used to predict stresses and deflections in jointed concrete airfield pavements serving the Boeing B-777 aircraft. The trained ANN model produces stresses and deflections with average errors less than 0.5 percent of those obtained directly from the finite element analyses. Use of the ANN model has been found to be very effective for correctly predicting ILLI-SLAB stresses and deflections, in less then a second, with no requirements of complicated finite element inputs. On the other hand, Elastic Layered Programs (ELPs) are currently being used in mechanistic-based pavement design procedures for the analysis of jointed concrete pavements. Corrections are required to such ELP solutions to account for the effects of finite slab size, load location on the slab, and load transfer efficiencies of the joints. This can be accomplished using the ANN model which is currently being expanded to handle all possible aircraft gear configurations with multiple-wheel loading conditions by the use of superposition principle. As demonstrated in this study for the solution of the B-777 aircraft gear loadings, trained neural network models will eventually enable pavement engineers to easily incorporate current sophisticated state-of-the-art technology into routine practical design.
    The important findings will be outlined in this oral presentation.

Speaker: Franz Rothlauf, Department of General Engineering
Time: 2:42pm
Title: A Program for the Design of Communication Networks
Author(s): Franz Rothlauf
Abstract: I will present a computer program for the design of communication networks. The software allows you to specify the neccessary amount of traffic between any two network locations. The cost function, which is implemented at the moment, calculates the cost of the resulting links based on the cost structure of the German telecom. But you can adjust it to any other cost structure, which depends on distance between network nodes and capacity of the line. By the help of either simulated annealing, cooperative simulated annealing or genetic algorithms the software calculates the cost minimized solution of the network. At the moment there are solutions implemented for tree networks and tree networks with one backbone.
 

Session III (5602 Beckman)

Speaker: Camille Goudeseune, School of Music and NCSA
Time: 2:55pm
Title: Using Genetic Algorithms to Optimize R^m-to-Rⁿ Mappings for Real-time Control
Author(s): Camille Goudeseune
Abstract: Sound synthesis algorithms often have several dozen control parameters. Humans cannot effectively control this many parameters simultaneously. We have developed a method of controlling a few dozen parameters with only two or three, essentially embedding a wrinkled slice of three-space in Rⁿ so as to pass through a priori chosen points in Rⁿ which produce desirable sounds. We use a genetic algorithm to find an optimal configuration of preimages of these points in R³.
    I will demonstrate an e-violin controlling real-time sound synthesis by means of its acoustic signal and its spatial position/orientation, using this data to drive the maps produced by the GA.

Speaker: Richard D. Braatz, Department of Chemical Engineering
Time: 3:05pm
Title: A General Framework for the Analysis and Control of Nonlinear Dynamical Systems Modelled by Dynamic ANNs
Author(s): Richard D. Braatz
Abstract: The presentation will summarize activities in ANN modeling and control conducted over the past five years. The results are of three types:

• Proofs that early any nonlinear system can be approximated with arbitrary accuracy with a dynamic artificial neural network (DANN), either of state space or input-output form. This includes processes with unstable zero dynamics, chaos, multiple steady-states, and quasi-periodic behavior. These results extend related work by Narendra and others.
• Computationally-efficient architecture-independent conditions for analyzing the stability and performance of systems of DANNs. Performance is measured in terms of an induced 2-norm between inputs and outputs, and all conditions are written in terms of linear matrix inequality optimization or feasibility problems. The results are mostly new, but build on the frameworks of Zames, Safonov, Lure, Lyaponuv, and other early researchers in control and nonlinear dynamical systems theory.
• Applications of the analysis conditions within an optimization-based formalism for the design of nonlinear controllers. This framework does not require most of the assumptions commonly made in other nonlinear controller design approaches, such as stable zero dynamics or perfect state information.

Speaker: Gary Dell, Department of Psychology
Time: 3:30pm
Title: Impaired Lexical Access in Speech Production
Author(s): Gary Dell, UIUC, and Dan Foygel, CMU
Abstract: Interactive spreading activation models of retrieval have provided good accounts of the error patterns that occur when we speak. Here, we show that reducing the connection weights in such models leads to patterns that match those made by aphasic speakers. Specifically, we can fit the data of 21 patients with two free parameters – the strength of lexical-semantic connections, and the strength of the lexical-phonological connections.

Speaker: Dan Roth, Department of Computer Science
Time: 3:40pm
Title: Cognitive Computation
Author(s): Dan Roth et al
Abstract: I will talk about some of the research done by the Cognitive Computation group. Specifically, I will mention theoretical work done on learning coherent concepts – multiple learning tasks with mutual compatibility constraints on their outcomes, and on the SNoW learning system.

Speaker: Jay Mittenthal, Department of Cell and Structural Biology
Time: 3:45pm
Title: Modeling the Evolution of Signaling Networks within Cells
Author(s): Jay Mittenthal and David E. Goldberg
Abstract: We ask why intracellular pathways for signal transduction are so long – why they have more proteins in sequence than the minimum number needed to perform their apparent functions. We plan to test two hypotheses: (1) Selection for a larger number of pathways favors longer pathways; and (2) Selection for a larger number of distinct pathways favors longer pathways that use families of proteins. Using modified genetic algorithms we will simulate the evolution of a population of cells, with each cell containing a network of signal transduction proteins, to see how the networks evolve under selection for a larger number of pathways. This project will refine methods for future work on the evolution of molecular networks.
 

Session IV (5602 Beckman)

The following six speakers were asked to prepare a commentary on various issues regarding the future of adaptive computation, especially scaling up approaches that they have worked with to the real world. Each responded with a short talk on recent results, challenge problems and open questions, and promising new technologies relating to their research.

Speaker: David E. Goldberg, Department of General Engineering
Time: 4:00pm
Title: Some Recent Results in Genetic Algorithm Competence and Efficiency
Abstract: The Illinois Genetic Algorithm Laboratory has studied the development of competent and efficient genetic algorithms. Competent GAs are those that solve hard problems quickly, reliably, and accurately. Efficient ones are those that do so even more speedily. This talk examines a potpourri of recent results in competence and efficiency, including very recent results in time utilization, evaluation relaxation, parallelization, and hybridization.

Speaker: Jamshid Ghaboussi, Department of Civil Engineering
Time: 4:09pm
Title: Challenges in Applying Biologically Inspired Soft Computing Methods to Intractable Engineering Problems
Abstract: Biologically inspired soft computing methods such as neural networks, genetic algorithm and fuzzy logic have inherent characteristics suitable for dealing with the difficult and intractable engineering problems. The full utilization of the potential of these methods in engineering requires new ways of formulating the engineering problems and new ways of viewing the problem solving capabilities of the soft computing methods. We will consider two broad class of engineering problems: engineering design and inverse problems in engineering. The challenges in applying the biologically inpired methods to these problems will be discussed briefly. Some of my research group's work will be presented and discussed.

Speaker: Sylvian R. Ray, Department of Computer Science
Time: 4:18pm
Title: Environmental Adaptation and Lifelong Learning
Abstract: Among the challenging technical issues for the near future, one of the obvious ones is the advantage to be gained by introducing environmentally adaptive features into all types of mechanisms.
    A longer range goal, which could drive (a major subfield of) the development of adaptive computation methods, is "Lifelong Learning", that is, auto-adaptive systems which grow slowly and stably in harmony with their environment while improving their ability to perform their goals.
    Applications and testbeds are not lacking, as I will point out.

Speaker: David Tcheng, National Center for Supercomputing Applications (NCSA)
Time: 4:27pm
Title: Scaling Up Knowledge Discovery Methodologies to Real-World Applications
Abstract: Recent research in adaptive computation for machine learning and optimization has led to advances in construction of knowledge-based systems for various idealized problem domains. An impasse, however, still remains in scaling these techniques up to truly massive databases (by current standards, those of terabyte scale and larger). I will discuss a framework for high-performance inductive learning that treats many of the "free parameters" in configuration of a learning system as a high-level optimization problem. Examples of such parameter optimization include tuning of performance constraints, feature construction and subset selection, and decomposition of modular problems. Multiple learning methods, models, and sources of data must often be integrated to effectively analyze a distributed database. The accessibility of neural, probabilistic, and genetic algorithms through a visualization and visual programming environment is also important to their scalability. I will conclude with a description of current research at the NCSA Automated Learning Group on scalable data mining and knowledge discovery in very large databases for real-world applications, such as fraud detection, prognostic decision support, control automation, and document categorization for information retrieval.
Slides: for this talk (Microsoft PowerPoint 97)

Speaker: Stephen E. Levinson, Department of Electrical and Computer Engineering and Beckman Institute (Human-Computer Intelligent Interaction Division)
Time: 4:36pm
Title: The Role of Sensorimotor Function, Associative Memory and Reinforcement Learning in an Autonomous Intelligent Robot
Abstract: We are testing experimentally a novel theory of cognition by building an autonomous intelligent robot. Our theory comprises three equally important principles. First, we believe that cognition, as manifest in humans, requires a well integrated sensorimotor periphery. For the purposes of this experiment sensorimotor function includes binaural audio, stereo video, tactile sense, and proprioceptive control of motion and manipulation of objects. We are trying to exploit the synergy intrinsic in the combined sensorimotor signals. This sensory fusion is essential for the development of a semantic representation of reality from which all other levels of linguistic structure derive. Second, for intelligent behavior, the widely accepted role of computation in the sense of Turing is of secondary significance to the primary mechanism of associative memory. We are designing several different architectures of such a memory. And, finally, the contents of the associative memory must be acquired by the interaction of the machine with the physical world in a reinforcement training regime. The reinforcement signal is an evaluation of only the success or failure of the robot's behavior in response to external stimuli. This signal comes from three sources: autonomous exploration by the robot, instruction of the robot by a teacher as to the success or failure of its behavior, and instruction of the robot by the teacher in the form of direct physical demonstration of the desired behavior (e. g. overhauling the robot's actuators). Such instruction must make no use of any supervised training based on pre-classified data. Nor may the robot use any pre-determined representation of concepts or algorithms. The result should be a robot, trained as described above, of sufficient complexity to be able to carry out simple navigation and object manipulation tasks in response to naturally spoken commands. The robot is to acquire language along with its other cognitive abilities in the course of its training.

Speaker: Jesse Reichler, Department of Computer Science
Time: 4:45pm
Title: How to Build a Brain: Leveraging Adaptive Mechanisms for High-Dimensional Autonomous Learning and Control
Author(s): Jesse Reichler and Clay Holroyd
Abstract: A framework for building large-scale brain-like controllers is described, grounded in an interpretation of such systems as efficient approximations of decision theoretic "optimal" controllers; we discuss how such an approach sheds light on the structure and function of the nervous system.
    In the biological brain, multiple hierarchies of perception, action, planning, and learning operate smoothly at a wide range of time-scales. We are interested in understanding computational principles of organization that support large-scale cooperative computation; our aim is to develop a simple, pedagogically clear example of a "complete" autonomous controller that demonstrates reasonable interaction between attentional, memory, motor, and perceptual systems, and exhibits adaptation on a range of time-scales. An important aspect of the work involves drawing connections between the controller architecture and biological structures in the brain, including cortex, cerebellum, basal ganglia.
    This talk will be a broad overview which focuses on the ways that various adaptive mechanisms can be leveraged to make efficient use of available resources.
    If time permits I will discuss areas of potential collaboration between researchers in AI, GAs, NNs, and control theory, and describe a robotic simulator which we have developed to perform control experiments.
 

Back to the ANNCBT main page