For the midterm project you will explore the GNG model as described in A Growing Neural Gas Network Learns Topologies (Fritke, 1995) and aspects of the SODA model as described in Developing Navigation Behavior Through Distinctive State Abstraction (Provost, Kuipers, and Mikkulainen, 2006).
Because time is limited for the midterm project, you should not attempt the reinforcement learning portion of the SODA model. Instead, you should focus on steps 2-5 described on page 7 of the Provost et. al. paper. The goal will be to generate a set of high-level actions that use the high-level perceptual features generated by the GNG. The high-level actions will consist of a trajectory following control law combined with a hill climbing control law. After implementing high-level actions in this way, you may also explore other possibilities.
You will design an experiment, run the experiment, analyze the results, and write a paper to explain what you found. You are strongly encouraged to work with a partner on this project. You may also work in teams of three, if the project is substantial enough.
Do an update81 to get a copy of my code that implements the basic framework of the GNG model. SODA uses a variant of GNG that only adds nodes when the average error reaches a threshold and also defines an activation function across the network. You will need to implement these new aspects.
You may design an experiment that uses one of the actual robots available in the lab (Aibos, Pioneer, or Khepera). You are also welcome to conduct the experiment entirely in simulation. Whether you use an actual robot or a simulated robot, you will need to carefully consider the environment. What actions and sensory information should be available to the robot? One option is to try to reproduce the original experiment as closely as possible. Another key question to consider is: How will you analyze the resulting high-level actions.
In addition, you are expected to conduct a limited literature review. Find and read at least two articles that are related to your experiment. A good source of related work are the bibliographies of the original papers.
Be ready to discuss your ideas for the project during our next class meeting.
Your grade will not be based on whether or not your experiment succeeds. Negative results are also useful. Your grade will be based on the design and execution of the experiment as well as the thoroughness and readability of the paper. You do not turn in any programs. Bring your completed paper to my office.
The purpose of writing a research paper is to communicate your discoveries to others. It is important to explicitly acknowledge how your work relates to other work. This can be done either with a direct quotation from another source or by summarizing the key points from another source. You should avoid paraphrasing another source as this can border on plagiarism. When summarizing another source, explain the essential information in your own words.
When using the exact language from another source, you must use quotation marks. Or if you are using a passage that is more than four lines long, indent and single space the passage without quotation marks. For a direct quotation, provide a reference with a page number. When summarizing another source, you should also provide a reference, but a page number is not necessary.