David J. Wu

From Chessprogramming wiki
Revision as of 19:32, 11 May 2019 by GerdIsenberg (talk | contribs)
Jump to: navigation, search

Home * People * David J. Wu

David J. Wu [1]

David Jian Wu,
an American computer scientist and computer games researcher and programmer, who balances a mixture of software development and research within the financial industry [2]. He defended his B.Sc. degree in 2011 at Harvard College, Harvard University, delivering the thesis Move Ranking and Evaluation in the Game of Arimaa [3]. David J. Wu is author of the Arimaa bot Sharp, and inspired by AlphaZero, the Go playing program KataGo .

Sharp

The Arimaa playing bot Sharp won the 2015 Arimaa Challenge and the then $12,000 USD prize by defeating each of three top-ranked human players in a three game series [4]. It already played the 2008 Arimaa computer tournament, and became runner-up behind David Fotland’s program Bomb, and further won the 2011 and 2014 tournaments but not the contest against the best human players of that time [5]. Sharp's design was elaborated by its author in the 2015 ICGA Journal, Vol. 38, No. 1 [6]. It follows the same fundamental design as strong Chess programs, using an iterative deepening depth limited alpha-beta search and various enhancements within a parallel search algorithm conceptually similar to the dynamic tree splitting described by Robert Hyatt in 1994 [7]. Sharp further implements several Arimaa-specific search enhancements with four steps per move, such as static goal detection and capture generation, and continues to use and benefit greatly from a move ordering function developed in 2011 as described in Wu's thesis - the move ordering function is the result of training a slightly generalized Bradley-Terry model over thousands of expert Arimaa games to learn to predict expert player's moves, using the same optimization procedure described by Rémi Coulom for computer Go [8].

KataGo

KataGo is a Go playing entity inspired by the AlphaZero approach of combining Deep learning with Monte-Carlo Tree Search (MCTS) using pure reinforcement learning aka self play to train the deep neural network. Due to modifications and enhancements of the AlphaZero-like training process, self-play with a only few strong GPUs of between one and several days is sufficient to reach somewhere in the range of strong-kyu up to mid-dan strength on the full 19x19 board [9].

Selected Publications

[10]

External Links

References

Up one level