CSCI323/MCS9323: Artificial Intelligence

PROJECT GUIDELINES

• The project is worth 30% of your final mark. The project is meant to be an exercise for you to do something creative with the techniques and technologies we are discussing in CSCI323. The project must involve practical applications of artificial intelligence, or theoretical insights/advances in the area. Your project should be designed to fit one of the following three categories:
  • Category One: You come up with an idea that makes an interesting and (at least somewhat) original contribution to the state-of-the-art. For instance, this could be a new variant of an existing knowledge representation language, a new variant of an existing search technique, a new agent programming language, or a new technique to support agent negotiation. You describe your ideas in a concise report of 10-15 pages (content matters, not size). If your idea is interesting/original enough, you don't need to write and experiment with code to prove validate your idea.
  • Category Two: You come up with a novel/original application of a technique that we have discussed in class. You describe your idea in the report, possibly including a system design for the proposed application domain. You write some code by way of a (at least partial) proof-of-concept implementation. Your implementation will not be expected to be as detailed and as comprehensive as in a Category Three project (described below). You include an explanation of the code in your report and attach the documented source code.
  • Category Three: You write code to implement and test some technique that we have studied in class (or is related to the course content) in a realistic/practical application. The code you write should be substantive (justifying 10 weeks of work). Your report will describe the implementation and explain why it is interesting. You will attach the documented source code to your report.
  Note that these categories are not presented in any specific order of importance . You may do well with a project in any category.
• The following are some interesting ideas for possible projects. Note that th is list is not meant to be complete.
  • Novel techniques for knowledge representation; Consider for instance the latest extension to logic programming that has lots of AI researchers excited - answer set programming (this page has some good initial pointers). Can you extend the knowledge representation language, can you develop better ways of computing answer sets, or can you think of novel applications of the technology>
  • Agent technology applications: The UMBC AgentWeb is a good collection of resources on these topics. Botspot lists all kinds of agent applications.
    • Look at how agent building tools (such as those listed here) and agent programming languages (such as AgentSpeak) work. Are there interesting applications for these tools and languages ? Can you suggest improvements to any of these ? Can y ou invent new ones ?
    • Look at how agent technology is applied in e-commerce (see this page for a good collection of introductory papers). Can you think of new ways of building agents for e-commerce applications ? Can you think of new e-commerce related application domains for agent technology ? Can you build or improve upon an agent-based marketplace such as Kasbah ?
    • Look at how agent technology is being used to solve resource allocation prob lems (this page introduces market-oriented programming. Can you think of new applications for such techniques ?
    • Look at agent trading applications, such as those listed in the Trading Agent Competition webpage.
    • Look at applications of multi-agent systems, such as the fun Robocup competition (or our very own Robocup team - Gongeroos'99 ) or the more serious Robocup Rescue competition.
    • Look at commercial ventures such as Agentis and Agent Oriented Software. Can you think of new and innovative applications along these lines ?
  • Constraint programming and intelligent search technology: Can you think of better ways of representing and solving constraint satisfaction problems, or real-life optimization problems? This page is a good collection of resources on constraint programming Begin by looking at this page for introductory material. This webpage is specially useful.
  • Image compression using AAN (auto-associative neural network): Propose a mean to simulate an AAN by using MLP with a single hidden layer. Use the MLP as a tool to learn the compressing and uncompressing of images. Note that the first layer in the network which connects the input with the hidden neurons can be seen as a compressor of data presented to the network, while the second layer which connects hidden neurons with the output neurons can be seen as an un-compressor. Run some experiments by using the dataset of images from http://www.uow.edu.au/~markus/teaching/CSCI323. What is the MLP configuration (ie. what is the number of hidden layer neurons, and what parameters did you try?) that allows for a reasonable compression of all images, while producing a performance with acceptable error? Sources with introductory material to MLP and AAN are plentiful. For example, introductory material on MLP can be found http://hebb.mit.edu/courses/9.641/2002/lectures/lecture04.pdf
  • Identification of appropriate training parameters in SOM (Self Organizing Maps): Current approaches to identify a suitable training parameters (eg. learning rate, neighbourhood radius, training iteration and weights) in SOM largely rely on trial-and-error. Propose an approach to identify appropriate parameters for training that avoids trial-and-error. Explain and show how your approach works, then represent your training result on a map. Some introductory material on SOM can be found in scholarpedia or in parts 1 & 2 from The Peltarion blog, or by looking at a book by T.Kohonen (in library). A good implementation (binary and source code) of the SOM can be found at http://www.cis.hut.fi/research/som_pak/ the software is explained in the following document http://www.cis.hut.fi/research/papers/som_tr96.ps.Z you may modify the above sources, use other sources, or implement a SOM simulator yourself.
  • SOM (Self Organizing Maps) for cyclic graphs: SOM has a limitation on dealing with cyclic graphs, and as a result, some pre-processing before training is usually required for such graphs. Read up on what has been done in the past (eg. CSOM-SD), then propose a novel approach that allows the processing of cyclic graphs by a SOM. Show that the computational complexity of your approach has been taken into account.