The Spades BidderProject OutlineBackgroundPreliminary WorkData GatheringModular Network ConstructionNil BidderTrump BidderNon-trump BidderResultsConclusionThe Spades BidderCS 539 Spring 2000Instructor: Dr Yu Hen HuBy: Jeremiah NeubertProject Outline•Background•Preliminary work•Data gathering•Bidder construction•Modular network•Components•Results•ConclusionBackground•4 players•13 cards each•Must follow suit•Spades are trump•Bid on how many books the hand can take•Each person bids once•Play with partners to take the tricks•Yahoo rules usedPreliminary Work•52 Input network•Divided network•Reduced input divided bidder•Results (Final Design)–3 experts–Less noise–More like human bidderData Gathering•Making hands–Distribute cards between four players–Format the hands to fit bidder•100 Examples per expert per bidder •50 Hands played to train modular networkModular Network Construction•3 experts•Nil bid–Voting system using nil bidder input–Traditional bid less then 4•Tradition bid–Sums input of non-trump and trump bidder–Combines the inputs of the 3 experts using varianceNil Bidder•MLP network–1 Hidden layer•Inputs–High cards in each suit–Number of low cards in each suit–Number of cards in each suit•One outputTrump Bidder•MLP network–2 hidden layers•Inputs–High spades–Number of cards in each suit•15 outputs–Each represents 1/2 BookNon-trump Bidder•MLP network–1 hidden layer•Called once for each non-trump suit•Inputs–High cards in the suit–Number of card in the suit•8 outputs–Each represents 1/2 bookResults•11 hands played•Human player•Robot partners and opponents•Achieved goal set–Robot bidder score -620–ANN bidder score +600Conclusion•Positives–Good results vs. current technology–Bids most hands well–Rarely overbids hands•Possible Future Work–Lack of rare hand data–No player feedback–Adjusting bid to fit
View Full Document
Unlocking...