end of era: human go supremacy
try to watch alphago movie
2006 start of computer go revolution
MCTS, Crazystone Coulom
UCB, UCT Kocsis+Szepesvari
UCT, patterns, MoGo Gelly+Wang+Munos+Teytaud
2015, top Go programs Zen, Crazystone, about 8dan
2014 Crazystone + 4 stones defeats Norimota Yoda by 2.5 points
weaker than top amateur
weaker than any pro (around 1200 go pros)
boom
2016 Jan 28 AlphaGo Nature paper
2015 Oct AlphaGo v Fan Hui 5-0
AG to play Lee Sedol March 2016 Seoul
AG-LS pre-game predictions
AG-FH match: many AG moves sub-optimal, throwing away points
top pro should easily beat this version of AG
so LS will win ?
but some sub-optimal moves were not errors
AG picks move that maximizes est. win-prob, not est. win-score,
and … between matches
AG algorithm changed
AG NNs trained non-stop (and improved)
2016 Mar AG v Sedol 4-1
game 1, LS tries unususal early moves to throw AG off its game, fails
10 Mar game 2 move 37 to me, era ends here (commentary Michael Redmond)
more on AG-Sedol and computer go
how does AlphaGo work ?
in this course, we explore basic algorithms for solving puzzles and games
general algorithmic principles: search, knowledge, simulations
by the end of the course you will learn most of the ideas behind AlphaGo
missing pieces (image recognition via deep convolution neural nets) in CMPUT 496
DCNN (not covered in this course)