Quoridor-Zero / README.md
rwr9857's picture
Create README.md
9556f4e verified
|
Raw History Blame Contribute Delete
3.96 kB
metadata
license: unknown

데이터셋 요약 (Quoridor AI Training Dataset)

이 데이터셋은 강화 학습 기반 Quoridor AI 훈련을 위한 Self-Play 게임 기록으로 구성되어 있습니다.
각 샘플은 게임 중 하나의 상태(state), 해당 상태에서의 정책(policy), 그리고 그 상태의 가치(value)를 포함합니다.


컬럼 설명

컬럼 이름 설명 데이터 타입
state 게임 상태를 나타내는 3차원 배열 list[list[list[int]]]
policies 각 행동에 대한 확률 분포 (정책), 총합은 1 list[float]
value 해당 상태의 승리 가치: 승리 1, 패배 -1, 무승부 0 int (1, -1, 0)

게임 진행 및 라벨링 기준

  • 승리한 에이전트의 모든 state 샘플에는 value = 1
  • 패배한 에이전트의 모든 state 샘플에는 value = -1
  • 무승부로 간주된 게임의 모든 state 샘플에는 value = 0
    (무승부 기준 : 100 Turn이 넘어도 승리한 에이전트가 존재하지 않을 때)

정책(policies)

  • 고정된 행동 수에 따른 확률 분포
    • quoridor-mini: 44개 행동
    • quoridor: 140개 행동
  • 각 값은 해당 행동이 선택될 확률이며, 총합은 항상 1

예시: [0.0, 0.0, 0.6, 0.0, 0.3, 0.0, 0.1, ...]
→ 세 번째, 다섯 번째, 일곱 번째 행동에 대해 확률 분포가 주어짐.


행동 인덱스 규칙 (일부 예시)

행동 방향 인덱스 번호
↑ up 0
→ right 2
↓ down 4
← left 6
... 기타 벽 설치 및 특수 행동 등

분석 코드 예시

from pathlib import Path
import pickle
import pandas as pd
import numpy as np


history_file = Path('./20240502082822-10K.history')

# history 파일 불러오기
with open(history_file, 'rb') as f:
    history_data = pickle.load(f)

# 데이터 프레임 생성
df = pd.DataFrame(history_data, columns = ['state','policies','value'])


'''
curr_player : 1
curr_wall : 3
oppo_player : 2
opppo_player : 4
'''
def convert_board(state):
    # quoridor size : 17
    # quoridor mini size : 9
    size = 17
    board = [[0 for _ in range(size)] for _ in range(size)]

    curr = np.array(state[0][0])
    curr = np.where(curr == 1)[0]

    currRow = (curr // size)[0]
    currCol = (curr % size)[0]
    board[currRow][currCol] = 1

    curr_Wall = np.array(state[0][1])
    curr_Wall = np.where(curr_Wall == 1)[0]
    for wall in curr_Wall:
        currRow = wall // size
        currCol = wall % size
        board[currRow][currCol] = 3

    oppo = np.array(state[1][0])
    oppo = np.where(oppo == 1)[0]

    oppoRow = (size - 1) - (oppo // size)[0]
    oppoCol = (size - 1) - (oppo % size)[0]
    board[oppoRow][oppoCol] = 2

    oppo_Wall = np.array(state[1][1])
    oppo_Wall = np.where(oppo_Wall == 1)[0]
    for wall in oppo_Wall:
        oppoRow = (size - 1) - wall // size
        oppoCol = (size - 1) - wall % size
        board[oppoRow][oppoCol] = 4

    return board

# n번째 행 가져오기
nth_row = 0

# state 표시
state = df.iloc[[nth_row],[0]] # DataFrame df의 n번째 행과 0번째 열에서 값을 선택
stateNP = state.to_numpy()[0][0] # DataFrame을 NumPy 배열로 변환
board = convert_board(stateNP)

# board 표시
board_np = np.array(board)
np.set_printoptions(linewidth=200)  # 줄바꿈 방지

policies = df.iloc[[nth_row],[1]].to_numpy()[0][0]
value = df.iloc[[nth_row],[2]].to_numpy()[0][0]

print('=== state ===')
print(board_np)
print('=== policies ===')
print(policies)
print('=== value ===')
print(value)