GAMES

axis

Who:

- Some people from the TAO team , in particular the tao-uct-sig
- Some people from the PRISM lab
- One person from the Math. lab of Paris-Sud
- Some collaborators in Taiwan (publications with NUTN, NDHU)

(there are intersections, so some items appear in several parts below...)

- Combining AIs: a simple tool for enhancing virtually any stochastic game AI by learning a random seed probability distribution.
- Tested games:
- 0-sum matrix games: this is a simple Octave solver (might be ok for Matlab as well, I did not check! no warranty, use at your own risk!)
- Starting! phantom go matrices under construction, for analysis. Great challenge from a fundamental point of view, and a nice test bed for all problems with adversarial and partially observable components.
- the Batoo game (simplified!) (creation - gnugo level 10)
- the batoo game with several games per metagaming choice (backup)
- the game of Go
- we start considering the Pokemon card game
- we also considered Urban Rival
- matrix of gnugo games with one row and one column per variant
- RS bootstrapping: matrix 1 for learning + matrix 1 for testing generalization (alpha-beta vs alpha-beta + alpha-beta vs other alpha-beta, both 9x9 komi default=6.5, random seeds); new version with monte carlo: learning, generalisation

- Bandit
- Decidability, complexity
- Undecidability of two-player partially observable games, even with finite state space and deterministic transitions

- Games, problems, ranking:
- Generating problems: tsumegos
- Deducing the strength of a player from his moves.

- MCTS
- Double progressive widening, theoretical analysis / consistency proof, submitted.
- Double progressive widening
- Application to Minesweeper, by combining myopic algorithms and MCTS
- Application to the game of Go
- Improving exploration

- Partial Observation
- Undecidability of two-player partially observable games, even with finite state space and deterministic transitions
- Application to Urban Rivals
- Double-tree MCTS

- Not directly part of this work, but related, we also like applications to energy, which have not replaced traditional games as our main applicative target