back to intelligent systems

So you want to publish a paper? Or perhaps you're interested in the latest developments in Intelligent Systems and Artificial Intelligence? Or maybe you want to know what goes on at the Genetic and Evolutionary Computation Conference (GECCO)? Whichever your interest, I will share with you my experiences in publishing and presenting a paper for LTU at GECCO 2002 as well as attending many of the presentations.

First, GECCO is organized and sponsored by both the International Society for Genetic and Evolutionary Computation (ISGEC) and the American Association for Artificial Intelligence (AAAI). The purpose of the conference as stated on their website (http://www-illigal.ge.uiuc.edu:8080/GECCO-2002/) is to "…present the latest high-quality results in the growing field of genetic and evolutionary computation." Topics presented included:

In order to publishing a paper, you first must have some original work to share that is of interest to the conference membership. In my case, this involved a practical application of evolutionary systems to determine the best schedules for courses in the Lawrence Math and Computer Science Department.

Determining the best, or near best timetable of lecture/courses for a university department, which optimizes enrollment, is a challenging problem. Described herein is a platform developed to accept student-generated data of course preferences, the university department curriculum, and resource constraints. The platform then generates the best or near best feasible schedule using an Evolutionary Computation Algorithm combining crossover and mutation for a fixed population. In addition, the platform permits manipulation of the population size, the initial mutation rate, the rule for varying the mutation rate, and the crossover point the platform order to examine the impact of these parameters and to determine the best values for this class of timetable problem. While this is an ongoing project, included is a sample run on an artificial data sample which shows the efficiency of the algorithm.

Next, you must determine at which conference you want publish. This is obviously based primarily on topic, but there are other considerations such as timing, location and journal affiliations. Timing is important, as submission deadlines are typically6 or more months in advance. Though in many cases only an abstract of your work is required. Location severely impacts the cost of attendance, which is required if your work is to be published. Finally, the organizations sponsoring the conference determine which journals will actually publish your paper.

In my case there wasn't much choice if I was to publish during the summer of 2002. GECCO 2002 was definitely in my field of interest, had a June 3 deadline for submission of "Late-Breaking Papers" and was located in New York City. Thus the deadline was possible to meet and the cost of attendance possible (compared to Conferences in Singapore or elsewhere in the world where travel costs alone would prevent my attendance).

Given that you have completed some work with findings that you can write about, the physical writing of the paper is not difficult. You do, however, have to pay careful attention to the conference's formatting rules. If papers are to be published they must conform to this format. This gets down to specific line spacing, font and font sizes for text versus code versus pseudo code, section and subsection headers and numbering, picture/graph labels and more. They also specifically required two separate first pages, one without author information to allow blind reviews of the paper by a panel of peers. Fortunately, GECCO provided MS Word Templates as well as PDF files of instructions for formatting and I am sure that is common for most conferences today.

Of course, once you submit your work, it is a waiting process to see if it is accepted. Similar to waiting to see what your GRE or SAT scores are, or if you have been accepted at the University of your choice!

The first questioner asked about the scope of our program or whether the system would generate an entire University Schedule or just the department schedule. In answer, the system is being implemented for the department on test basis only at this time. However, all indications are that the system would be able to handle a University's complete catalogue of courses and student input. With very large inputs, it might take several days to generate a solution, but that is reasonable for this type of problem.

The second questioner was concerned with the method used for encoding possible schedules for consideration. He felt that the input might limit the search from discovering valid answers. He was unable to describe the precise nature of his concern, even "offline" after our session and offered to read my paper and send a more precise question/comment in the future.

Part of his concern might have been with the encoding data in character strings rather than binary strings, as is usually done to generate individuals in GA's. This leads to the concern that my crossover method might not be truly "random" in nature and avoid some potentially unique solutions.

The conference itself started with workshops the first day. There was a full day
workshop for graduate students who wanted to present their work to a panel of mentors who critiqued the work and suggested future studies. Attendance to this workshop was limited and required submission by February.

The rest of the workshops were each 4 hours long, so it was possible to attend two out of 11 possible workshops. I particularly enjoyed the workshop run by Ian Parmee "Toward Interactive Evolutionary Search and Exploration Systems." This workshop focused on the current and future applications of Interactive Evolutionary Search and Exploration Systems.

To date, most of applications have been focused on developing artwork or music. In a sense, the method is similar to running a series of searches on a search engine such as Yahoo where you iteratively review the generated results and adjust/select new search criteria. You may have a sense of what you are looking for, but you don't know exactly what it is. As you narrow the search you get closer and closer to your ideal until the found solution is satisfactory.

In the case of music or artwork, the result is similar. The evolutionary search and exploration system establishes a base population of generated "art" based on initial parameters. The user indicates ranks the individuals in the population in terms of which is perceived to be better or a fit individual, and the system generates several more generations using those criteria to evolve better "art". The process then is repeated until the user is satisfied (or gives up from exhaustion).

Exhaustion is a very real factor. We as humans can be patient for only so long. An interactive system requires regular input for selection. Genetic Algorithms (GA's) and Evolutionary Systems (ES's) often require hundreds or more generations before a satisfactory solution is found! If human input is being used, the system then needs to find a way to characterize that input and use it for intermediate generations.

Current and future applications of Interactive Systems include use of these systems in engineering design. Ian Parmee's current work is focused on using such a system to locate fruitful areas of design for aircraft by specifying ranges of characteristics that impact airspeed, cargo capacity, maneuverability, etc. Another gentleman was working on developing a system that would take musical input and generate accompanying rythms and melodies in real time, rather like a jazz jam session.

The next day began with a series of four 1-1/2 hour tutorials. Once again, there were eight different programs to choose from in any given time-slot, so there was always something of interest to see and it was impossible to participate in everything. One presentation in particular was of interest to me that day, involving Evolutionary Robotics.

This presentation was made by A. Schultz and M. Potter of the U.S. Navy and was quite fascinating. They covered the traditional explanation of evolutionary systems quite rapidly and then went on to focus on practical aspects of implementing evolutionary systems in robotics. Two issues in particular were presented with examples.

First, there is the issue of "good visual sensors". The simplest and least expensive sensors are typically only capable of sensing light intensity (light and dark) in a given direction. These can be used in following dark lines along a white background, or locating a dark image or object in a surrounding light colored environment. Sophisticated image recognition is not possible. A step up might be some form of digital camera. However, the raw data requires a fair amount of computational power to analyse and generate into data. In a real-time system, this could require a separate CPU or a sensor with built-in analysis hardware. Another step up would provide a measure of distance or range to objects. An 2 dimensional image has no precise information in the 3^rd dimension though we might infer some relative relationships based on image overlapping, etc. Use of detectors that implement binocular vision with triangulation to determine distance or other method is essential if a robot is to navigate succesfully.

Another issue is modeling the environment in which the robot will operate accurately and, knowing it is impossible to consider every possibility, retain a failsafe control mode for emergencies!! In one project, an evolutionary system was used to develop an autonomously flown full scale aircraft. The aircraft had flown successfully in virtual environments and also with limited autonomy in the real world. In its first truly autonomous test flight, however, it crashed on take-off.

In retrospect, it turns out that the seams between concrete sections in the runway caused the aircraft to bounce during take-off. This motion input to the aircraft sensors fooled the system into acting as if it were already in flight in turbulent conditions, which was not appropriate for take-off. The point is, this simple harmonic motion caused by the runway was not considered in the virtual environments, and there was no way to remotely switch to manual control once the aircraft "brain" lost control.

Why evolutionary systems for these application you may ask? There are numerous potential advantages. These systems can adapt to unforeseen circumstances such as malfunctioning sensors or mechanical systems or changes in environment (though adaptation can be slow which may not be suitable for real-time applications such as flying an aircraft). Also, the system may find a solution that is much better than our intuitively designed solution as measured in terms of efficiency, reliability, etc.

GECCO 2002 - Key Note Speaker David Leinweber

David Leinweber of Caltech gave a talk on Evolutionary Algorithms in Financial Markets. I wish I could say that as a result I will be retiring next week and, no, I can't share any of the secrets learned in his presentation. Unfortunately, nothing in life is that simple or easy.

Dr. Leinweber has been involved in developing and implementing these systems since the early 1980's, however, and was able to share some general observations of what does not work or what to avoid. For example, the more people examine market movements and try to model them on "key indicators" the more amazing the results found. Dr. Leinweber has data that shows that the world markets could be predicted with better than 99% accuracy over a several year span by using the prices of Butter and Cheese as well as sheep populations in Bangladesh!

Just because the market behavior matches the fluctuations of sheep populations in Bangladesh for the last two years doesn't mean that that is a good method of predicting future behavior. Similarily, models developed based on past market performance may not be useful for prediction, particularly as time passes.

The point is that in a truly efficient market (if you accept that model) prices reflect all public information accurately. In other words, stock prices are primarily driven by current knowledge of the businesses and market conditions. Influences of past performance are not strong. For example, two investors see that the price of XYZ Corp had dropped steadily for the last two weeks. One investor decides to cut his losses and sell, the other investor decides the price has bottomed and is going to go back up and buys. The actions of the two investors cancel each other out in terms of market price.

So, does that mean that evolutionary algorithms can't be successful in the market? Not at all! Evolutionary systems can be used in data mining to get more current and accurate information before the rest of the world. Also, they can develop models much faster than humans can and provide insight for the human investor in making decisions. The key lies in informational advantages in every case. Like in Football where a tight-end fakes out and beats player covering him to catch the pass and make a great play, the investor can move ahead of the rest of the market given the right instincts and information.

Laura Landweber of Princeton University spoke about her work in developing Genetic Computers. In this case, she pointed to how actual life-forms such as single cell animals "compute" using DNA and RNA and went on to describe experiments she has performed in which RNA is used to solve problems in the test tube!

By encoding the RNA to represent positions on the board and implementing restrictions on sequences (to prevent two pieces being in check) it is possible to have a solution of RNA combine and recombine to find optimal solutions. In essence, this genetic material follows an ES process building populations of strands of RNA. While the 3x3 board above is quite simple, the number of legal solutions for larger boards up to 8x8 is huge. Even maximized, there are many more than one solution for these more complex problems.

The challenge at this time is in setting up the experiment. Biochemical methods are used to control how the encoded RNA strings are combined, and what makes for a survivor or more fit solution. Then, after the population has generated solutions, extracting the data is an intensive process of reading several RNA sequences and translating them back to the original problem syntax.

At this time, this methodology is confined to the laboratory and requires considerable time and effort to set up. The ability to solve extraordinarily complex problems rapidly using very small volume processors may have some unique advantages over traditional computer technology in the future.

At the end of the day, I believe that attending conferences such as these is extremely worthwhile. There is so much new research and so many new applications being developed it is impossible to keep up with them all. In the conference environment, however, you have an opportunity to hear the latest about your specific interests in a formal setting as well as to interact with professionals and academics from around the world and learn about their perceptions and interests.

Publishing and presenting your own work, also goes a long way to making attendance at a conference even more rewarding. Because of the scrutiny you have to place on you own work in initially publishing a paper, you become much more aware of the myriad of alternate solutions and methods possible for your problem. So in the conference, when you hear a new idea (or at least new to you) or field a question from the audience, it drives a whole new list of concepts and ideas for future work.

Finally, representing your university or employer in such a setting is an honor. Not only is it your name and work being reviewed and evaluated, it is an opportunity to let the world know about Lawrence Technological University and what the faculty and students are achieving through theory and practice! While that may sound a little corny to those of us who hear our school motto on a regular basis, there is no doubt that putting theory into practice teaches excellence and drives innovation.

back to intelligent systems