TD-lambda versus Learning by Search: Empirical study of two learning algorithms
contact: 
Let me say at the beginning, that this projet has a drawback within the Group of Microelectronics, being that it has nothing to do with microelectronics what so ever. It is purely algorithmic and will involve quite a bit of C and Java programming.
Introduction
A few years ago i conceived an algorithm to train a (classic error backpropagation) neural network for games. I called it 'Learning by Search' (LS). Another algorithm that has been applied to the same problem and that has been tremendously successfull is called TD-lambda. I consider LS to be theoretically better founded and more specialized in games, whereas TD-lambda is more heuristic but more ganerally applicable to real world problems.
I conducted a number of experiments, using LS to train a computer-player for Backgammon. The results are very satisfying, but a systematic comparison to TD-lambda has not yet been made. On my homepage, one can play against one of the best players I was able to train.
The Project
The project would be to perform such a systematic comparison between the two algorithms. I thought to do this in the form of a duel. Both algorithms can learn as they play games. TD-lambda can be said to learn from the game, whereas LS learns rather from contemplating the positions that occure during the game. The two algorithms could be made to play each other and their progress could be measured against each other.
Required skills
Well structured programming within a rather extensive piece of code will be required. Knowledge in both Java and C is necessary. Much of the code does already exist. The learning part of it will have to be adapted to run on todays system however, since it had been programmed a while ago for a special parallel machine. The TD-lambda algorithm will have to be included.
Some skills with mathematics/analysis will help to grasp the theory involved.