Please enable JavaScript.
Coggle requires JavaScript to display documents.
AlphaGo (Game of Go (Rules (Game of perfec information), Number of…
AlphaGo
Game of Go
Rules
Game of perfec information
Number of combinations
10³⁶⁰ different games
how does it work?
Ways of estimating
Monte Carlo tree search
MCTS
how works?
heurystyka
link Evernote
Zdolność wykrywania nowych faktów i związków między faktami
analiza najbardziej obiecujących ruchów
random sampling
in AplhaGo application
many games played to the very end
selecting moves at random
results of each game
weights of nodes used for future games
simulation's result
strong-amateur level
other usecases
other games
Pacman
Osadnicy z Cathanu
Magic: The Gathering
addition enhancement policies
predicting human expert moves
convolutional neural network (CNN)
goals of neural network
effective position evaluation (value network)
estimating probability that current move leads to win
action sampling (policy network)
training
stages of machine learning
Supervised Learning (SL)
trained from 30 million positions of
KGS
server
1 more item...
predict expert moves with accuracy 57%
slower evaluation
Fast Rollout (FR) policy
2 more items...
good in predicting next most likely moves
Reinforcement Learning (RL)
better in predicting best possible (winning) moves
1.2milion games
1 more item...
win 80% games against SL
win 85% with Pachi
1 more item...
Games with humans
with whom
Fan Hui
Result 5/0 for AG
Le Sedol
Źródło
machinelearnings.co