廢廢今天和專題老師討論過後,發現不至於使用三維矩陣,因為如同DAY15的作法,一個資料變為一個二維矩陣會變成169大小的矩陣只有一個數值而已,剩下的填0會有168為0的資料,會造成過度擬合,也就是資料重合率過高訓練出的結果沒有意義。
如圖,這是我所10次訓練出來的結果,準確度有100%,百分之百是過度擬合的問題啊:
因此我和老師討論出,換成15x15或是13x13(0110合併)的格式並且使用flop0~flop4的順序,
[strength, flop0, flop1, flop2, flop3, flop4, hand0, hand1, hands_level, player1_chips, player2_chips, 0, 1, 1, 0]
範例輸入:[12, 50, 38, 36, -1, -1, 32, 47, 4, 78.9, 39.0, 0, 1, 1, 0]
目前先2~A繼續用,並且使用10、20、30、40代表花色(梅花、方塊、紅心、黑桃)
/ | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | J | Q | K | A
------------- | -------------
flop0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 30
flop1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 30 | 0 | 0 | 0
flop2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 0 | 0 | 0
flop3 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
flop4 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
hand0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 0 | 0 | 0 | 0
hand1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 40 | 0
strength | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 12
hands_level | 0 | 0 | 0 | 0 | 4 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
player1_chips | 78.9 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
player2_chips | 0 | 39.0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0
現在廢廢要準備去,寫轉格式了!專題有進度了開心