(last updated 1998)

B.1 Examples of Projects in progress or to be Undertaken, 
                  the Research Component:

              What follows describes both examples of current projects as 
well as long-standing lines of research (themes) that may, of course, be 
redirected over time. A very good potential exists for funding of many of 
these projects by external foundations and industry/commercial sources. In 
fact, some of these projects have attracted initiation of research grants 
by national organizations such as ASME. Preliminary results are promising 
and external funds for extension and completion of the projects are 
actively pursued. Several  successful and ongoing relationships with local 
industry such as CAT, State Farm, Country Companies, Deere, etc. , have 
been developed.


Intelligent Distance Tutoring Systems, Intelligent Agents, Internet, 
Internet II, The World Wide Web, Intelligent Information Retrieval 
Project

The advent of the World Wide Web made available a huge amount of 
information on-line in the form of text and multi-media documents. The 
vast volume of information makes the traditional retrieval techniques, 
which are based on syntactic indexing inadequate. The problems of semantic 
information retrieval, automated conversion of information to knowledge 
and building domain knowledge bases can be effectively solved using AI 
techniques. The building of intelligent browsers and the application of 
Intelligent Agents technology to help users access and manage the 
available wealth of on-line information is also under study in this 
project.
               Another related goal of this project is to improve distance 
learning technologies by using the application of AI to the World Wide 
Web. With the possible participation of Bradley in the Internet II project 
and the availability of high speed networks for transmission of 
multi-media data, a lot of possibilities exist for application of AI 
technologies into constructing intelligent tutoring systems or smart 
courses. Real time audio and video availability, over the high bandwidth 
of Internet II, will certainly empower and make very applicable and 
desirable some of the new technologies to educators. Many problems exist 
in that context which are amenable to AI solutions. 
As part of this project, an on-line, web based, distance learning version 
of the curriculum for the certificate program and the concentration will 
be created. This will enable the certificate program to establish a global 
market base. In addition, it will provide a test-bed for the application 
of many of the AI algorithms and research in the intelligent tutoring 
systems and information retrieval areas. For example, student interactions 
with the on-line courseware can be recorded. Machine learning techniques 
from AI can be applied to customize the course (by recommending a sequence 
of course modules to be taken for example) and tailor it to the specific 
needs of each individual student. 

Neural Net Based Machinability Evaluation  

                The aim of this project is to conduct a systematic 
investigation of the machining process with the objective of developing a 
short term machinability test. Typically, it takes at least a week of 
testing by a skilled machinist to complete a material's machinability 
testing. In modern, global competition environments, this is very 
inefficient and uneconomical. Due to non-linear and complex relationships 
among the involved parameters, such as speed, feed, depth of cut, hardness 
of the material, etc., an analytical model is not known to exist to 
address the problem of machinability of materials.  In this project, 
neural networks are being used to address the problem of machinability. 
The feed-forward neural network model, trained with back-propagation is 
used for predicting the tool life during machinability. Preliminary 
results are promising. The model, if successful, will greatly reduce the 
time and expenses involved in machinability testing in industry. It has 
the potential to make a significant impact on the competitiveness of US 
industry.

Neural Net Project on Forging Die Design

One of the most important steps in controlling product quality during the 
metal-forming processes is the die and process sequence design. The 
ability to quickly change die design for new parts has become one of the 
critical manufacturing challenges. Unfortunately, there is a lack of a 
theoretical understanding of how the appropriate pre-form die shapes are 
achieved and what the exact effect of a chosen die design is on the formed 
billet. So, typically the design of a pre-form die shape is ad-hoc and 
usually based on human intuition and experience. In the case of studying 
the effect of a chosen design to the form of the resulting billet several 
finite-elements analysis based models have been proposed and commercial 
software is available which successfully predict various characteristics 
of the formed shape billet. This process is extremely computationally 
extensive and expensive. One could possibly use such simulation models to 
find the appropriate die design which would result in a desired billet 
shape, by the propose and test approach: propose a design; predict the 
shape of the formed billet using finite element analysis; repeat, until 
the difference between the formed billet and the proposed one is within an 
acceptable error range (threshold). But such an approach is impractical 
for computational and time constraint reasons. 
The objective of this research is to propose a reliable, efficient model 
for automating the forging pre-form die design process. Due to lack of 
underlying theory and an analytical model, the problem of optimizing the 
bust die design and the initial billet size to derive the desired forged 
piece is an ideal candidate for solving by using neural networks.

Hybrid Case Based Reasoning and Neural Network Design

Case Based Reasoning (CBR) is a breakthrough technology for knowledge 
based systems. The basic premise of this approach is that certain problems 
can be solved by retrieving relevant experience from previous similar 
situations. Early knowledge based systems had trouble improving their 
performance incrementally, in part because they could not remember or use 
the problems they had previously solved. CBR recognizes the need to use 
previous experience to build systems that improve over time. The CBR 
approach is to build a case base of previously solved problems. These 
cases have to be indexed using sophisticated retrieval methods. When a new 
problem is presented to the CBR system, a search is conducted through the 
case base to find the most similar problem, whose solution is then adapted 
to solve the current problem. 
	In this project, a CBR system is integrated with a neural network 
layer, which will enable the automated extraction of features or indexes 
from the given case base.  The problem of index construction in CBR 
systems is one of the major bottlenecks in their development. Automation 
of this task will greatly enhance the applicability of such systems. The 
integrated, hybrid system will then be applied to the domain of VLSI 
design.	

Autonomous Mobile Robots Project
 
Over the Fall of 1998 offering of CS 521, Introduction to Artificial 
Intelligence, class, several teams of students, both undergraduate and 
graduate, engaged in the design and building of micro-controller 
controlled mobile robots and implemented various algorithms on them, [1]. 
Mainly, contour following and obstacle avoidance algorithms were 
implemented. 
It was of great benefit to the success of the project to have team members 
with varied backgrounds, including computer science, electrical and 
mechanical engineering, and computer information systems. The completed 
intelligent system models the performance of an modern automobile with a 
driver, without the driver. Using this design, the lessons and methods 
learned from such an experiment could have many applications. For example, 
industrial robots could be programmed to transport materials.
This project will continue with an investigation of the use of neural 
networks and fuzzy logic to control the robots. The application of 
Minsky's Multi-Agent Intelligent Systems Approach, and Brooks' 
Subsumption-Based Methodology will both be explored and applied to a 
variety of robotic tasks. A set of algorithms will be developed to address 
obstacle-avoidance problems. Various techniques, including neural 
networks, will be employed for real-time determination of a robot's 
traveling route and speed. Various neural network models including a 
feed-forward, neural network, trained with back propagation, will be used 
to solve contour-following and obstacle-avoidance problems. 
 

     Other projects planed in this area include:
    
Communicating Agents/Artificial Life

Wireless communication will enable robots to talk to each other. 
Communities of communicating and collaborating robots will be designed and 
implemented. We envision many real-life commercial and industrial, as well 
as research, applications.

Autonomous Mobile Robot Navigation using Fuzzy Logic

               In this project, fuzzy logic is used to guide the 
navigation of an autonomous mobile robot system. Both contour-following 
and wall-following algorithms have been implemented using fuzzy, 
rule-based controllers. An evolutionary approach is introduced for 
designing membership functions for the various linguistic variables of the 
fuzzy system.
Design and specification of the membership functions of the fuzzy sets in 
a rule-based, fuzzy controller is a an ad-hoc process. It is usually 
achieved in consultation with a human expert from the particular domain of 
expertise. It has been observed that the performance of the fuzzy 
controller is greatly affected by the appropriate definitions of the 
various membership functions.
Genetic algorithms, ([3]), is a powerful optimization technique. It will 
be used for modification and optimization of the membership functions. 
From an original, reasonable approximation of the various membership 
functions, a population of individual intelligent systems, as defined by 
their membership functions, is constructed. This population can then 
evolve into optimized versions of the membership functions.