MOHEX Wins Hex Tournament
Ryan Hayward · ICGA Journal · 2012
PANORAMEX — named after Panoramix, the druid character from the Asterix and Obelix comic strip — uses the RAVE UCT formula (Gelly and Silver, 2007) with UCB exploration constant 0 and the save-bridge pattern in simulations. PANORAMEX ran on an 18-node cluster of 4-core machines, using root parallelization and majority vote to select each move. This yielded about 6×10 5 simulations per second. WOLVE, the 2010 silver medallist (Arneson, Hayward, and Henderson, 2010) 2, uses truncated-width alpha-beta search, a Shannon-style electric circuit evaluation function with cell adjacencies augmented by virtual connections, and pruning of inferior cells. To save time, WOLVE uses a book built by caching 6-ply moves. This year Broderick Arneson added pondering and changed the search algorithm from fixed-ply to variable-ply with timemanaged iterative deepening. WOLVE used 2 threads (one to select moves, one to solve) on a 4-core machine, reaching 6-ply on most moves. MOHEX, the 2010 gold medallist (Arneson et al., 2010), is a Monte Carlo tree search program built on the code base of FUEGO, the Go program developed by Martin Müller, Markus Enzenberger and others at the University of Alberta. FUEGO uses lock-free parallelization (Enzenberger and Müller, 2009), and backs up virtual losses for better parallelization. MOHEX computes virtual connections and inferior cells in UCT tree nodes visited at least 400 times. This year Arneson added pondering, Huang helped with tuning, and Pawlewicz improved the