[gnugo-devel] WinHonte and AI

[Måns Ullerstam]

I am sorry if I offended anyone regarding WinHonte. I was only trying to
make a point of testing some interesting machine learning techniques in
GNU Go. My plan is to try out a few different machine learning
algorithms in GNU Go and I was hoping that someone else on this list was
interested in the same area.

These are some of the approaches I am interesting in making:
- Automatically generate local patterns from human expert play, thereby
learning locally what to play and learning good shape as well.
- Automatically generate opening strategies (fuseki) reaching a lot
further than 5 moves (more or less the limit in the fuseki libraries I
have produced before with extract_fuseki)
- Try to use standard neural networks as a compact storage and retrieval
facility for patterns. For example you can train one network for each of
the pattern databases already included in GNU Go. In that case GNU Go
would be the teacher, telling the neural network how much all patterns
are worth and what to play in different situations
- Use TD(lambda), reinforcement learning, on different functions in the
game. That could be influence, score calculation, etc.

The main idea is that neural networks might take a long time to learn,
but are supposed to be rather fast at lookup. They are also very compact
in storage. It is also possible to capture more vague concepts and
generalize the knowledge, so it can be applied to more situations.

For example it might be possible for a neural network to learn what
moves or regions that are interesting to play on a full board scale.
Using this as a preprocessing phase to reduce the number of moves
evaluated by all other move generators.

I love the work you have done so far and my goal is to start
contributing again a lot more. I think it is a pity that there are so
many other small projects around the globe trying to solve different
issues with go, developing their own programs and user interfaces, etc,
instead of using GNU Go as the foundation and instead add a move
generator to GNU Go with their favourite theory applied. I want to
encourage people to work with GNU Go instead.

Best regards / Mans