概要: README.md
In [1]:
#-*- encoding: utf-8 -*-
%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
pd.set_option('display.precision', 4)
import sys
sys.path.append('./user_strategies')
import play as pl
from Iida_perfect_monitoring import Iida_pm
from Iida_imperfect_public import Iida_ipm
from Iida_imperfect_private import Iida_iprm
from kato import KatoStrategy
from ikegami_perfect import Self_Centered_perfect
from ikegami_imperfect_public import Self_Centered_public
from ikegami_imperfect_private import Self_Centered_private
from mhanami_Public_Strategy import PubStrategy
from mhanami_Imperfect_Public_Strategy import ImPubStrategy
from mhanami_Imperfect_Private_Strategy import ImPrivStrategy
from tsuyoshi import GrimTrigger
from gistfile1 import MyStrategy
from beeleb_Strategy import beeleb
from oyama import OyamaPerfectMonitoring, OyamaImperfectPublicMonitoring, OyamaImperfectPrivateMonitoring
from ogawa import ogawa
from yamagishi_impd import yamagishi
In [3]:
import unittest
class TestStrategies(unittest.TestCase):
def setUp(self):
self.Strategies = [Iida_pm, Iida_ipm, Iida_iprm, KatoStrategy, Self_Centered_perfect, \
Self_Centered_public, Self_Centered_private, PubStrategy, ImPubStrategy, ImPrivStrategy, \
MyStrategy, beeleb, OyamaPerfectMonitoring, \
OyamaImperfectPublicMonitoring, OyamaImperfectPrivateMonitoring, \
ogawa, yamagishi, GrimTrigger] # ここに自作のclassを入れる
self.case1 = "Signal is empty(period 1)"
self.case2 = [0, 1]
self.case3 = [1, 0]
self.case4 = [0, 1, 0, 1, 0, 0, 1]
self.seed = 11451
self.RandomState = np.random.RandomState(self.seed)
# case1を引数に渡してテスト
def test1(self):
print("testcase:", self.case1)
for Strategy in self.Strategies:
rst = Strategy(self.RandomState).play()
self.assertIsNotNone(rst, Strategy.__module__)
self.assertIn(rst, (0, 1), Strategy.__module__)
# case2を引数に渡してテスト
def test2(self):
print("testcase:", self.case2)
for Strategy in self.Strategies:
S = Strategy(self.RandomState)
for signal in self.case2:
S.get_signal(signal)
rst = S.play()
self.assertIsNotNone(rst, Strategy.__module__)
self.assertIn(rst, (0, 1), Strategy.__module__)
# case3を引数に渡してテスト
def test3(self):
print("testcase:", self.case3)
for Strategy in self.Strategies:
S = Strategy(self.RandomState)
for signal in self.case3:
S.get_signal(signal)
rst = S.play()
self.assertIsNotNone(rst, S.__module__)
self.assertIn(rst, (0, 1), S.__module__)
# case4を引数に渡してテスト
def test4(self):
print("testcase:", self.case4)
for Strategy in self.Strategies:
S = Strategy(self.RandomState)
for signal in self.case4:
S.get_signal(signal)
rst = S.play()
self.assertIsNotNone(rst, S.__module__)
self.assertIn(rst, (0, 1), S.__module__)
In [4]:
suite = unittest.TestLoader().loadTestsFromTestCase(TestStrategies)
unittest.TextTestRunner().run(suite)
Out[4]:
Test: OK
In [2]:
payoff = np.array([[4, 0], [5, 2]])
seed = 20141019
rs = np.random.RandomState(seed)
discount_v = 0.97
repeat = 1000
ts_length = rs.geometric(p=1-discount_v, size=1000) + 1
In [6]:
strategies = [Iida_pm, PubStrategy, KatoStrategy, Self_Centered_perfect,
GrimTrigger, MyStrategy, beeleb, OyamaPerfectMonitoring, ogawa,
yamagishi]
game = pl.RepeatedMatrixGame(payoff, strategies, signal=None, ts_length=ts_length, repeat=1000)
game.play(mtype="perfect", random_seed=seed, record=False)
In [7]:
score_perfect = np.array([[0, 96040, 60464, 147123, 124298, 147123, 147123, 95809, 106699, 126331],
[104995, 0, 68666, 137332, 137332, 137332, 137332, 137332, 97073, 137332],
[85969, 73666, 0, 170665, 122335, 104554, 82174, 73666, 76420, 73666],
[98168, 137332, 4000, 0, 137332, 137332, 137332, 137332, 83376, 137332],
[88338, 137332, 36220, 137332, 0, 137332, 137332, 137332, 93425, 137332],
[98168, 137332, 48074, 137332, 137332, 0, 137332, 137332, 99883, 137332],
[98168, 137332, 62994, 137332, 137332, 137332, 0, 137332, 105660, 137332],
[104604, 137332, 68666, 137332, 137332, 137332, 137332, 0, 96651, 137332],
[96454, 95428, 66830, 150821, 126385, 125958, 117485, 94806, 0, 110390],
[125791, 137332, 68666, 137332, 137332, 137332, 137332, 137332, 108350, 0]], dtype=float)
score_perfect /= 34333
score_perfect = pd.DataFrame(score_perfect)
print(score_perfect)
| Iida_pm | PubStrategy | KatoStrategy | Self_Centered_perfect | GrimTrigger | MyStrategy | beeleb | OyamaImperfectPerfectMonitoring | ogawa | yamagishi | |
|---|---|---|---|---|---|---|---|---|---|---|
| Iida_pm | 0.000 | 2.797 | 1.761 | 4.285 | 3.620 | 4.285 | 4.285 | 2.791 | 3.108 | 3.680 |
| PubStrategy | 3.058 | 0.000 | 2.000 | 4.000 | 4.000 | 4.000 | 4.000 | 4.000 | 2.827 | 4.000 |
| KatoStrategy | 2.504 | 2.146 | 0.000 | 4.971 | 3.563 | 3.045 | 2.393 | 2.146 | 2.226 | 2.146 |
| Self_Centered_perfect | 2.859 | 4.000 | 0.117 | 0.000 | 4.000 | 4.000 | 4.000 | 4.000 | 2.428 | 4.000 |
| GrimTrigger | 2.573 | 4.000 | 1.055 | 4.000 | 0.000 | 4.000 | 4.000 | 4.000 | 2.721 | 4.000 |
| MyStrategy | 2.859 | 4.000 | 1.400 | 4.000 | 4.000 | 0.000 | 4.000 | 4.000 | 2.909 | 4.000 |
| beeleb | 2.859 | 4.000 | 1.835 | 4.000 | 4.000 | 4.000 | 0.000 | 4.000 | 3.078 | 4.000 |
| OyamaPerfectMonitoring | 3.047 | 4.000 | 2.000 | 4.000 | 4.000 | 4.000 | 4.000 | 0.000 | 2.815 | 4.000 |
| ogawa | 2.809 | 2.779 | 1.947 | 4.393 | 3.681 | 3.669 | 3.422 | 2.761 | 0.000 | 3.215 |
| yamagishi | 3.664 | 4.000 | 2.000 | 4.000 | 4.000 | 4.000 | 4.000 | 4.000 | 3.156 | 0.000 |
In [8]:
def normalize(array_2d):
copy_array = np.array(array_2d, dtype=float)
for v in copy_array:
total = np.sum(v)
v /= total
return copy_array
def __plotC__(periods, action1, action2, **kwargs):
period_max = kwargs.get('period_max', np.max(periods)+1)
title = kwargs.get('title', None)
fig, ax = plt.subplots(figsize=(20, 5))
plt.xlim(-1, period_max)
plt.ylim(-0.1, 1.1)
plt.plot(periods, action1, color='c', label='player1 Cooperate %')
plt.plot(periods, action2, color='r', label='player2 Cooperate %')
if title:
plt.title(title)
plt.legend()
plt.show()
# period毎にどの程度協力が行われているかをplot
def plotC(df):
# 対戦組み合わせの数
match_len = df['Match'].max() + 1
# 1対戦のゲーム数
round_len = df['Round'].max() + 1
# 最長period
period_max = df['Period'].max() + 1
for m in range(match_len):
df_match = df[df_perfect.Match == m].reset_index()
str1 = df_match['Strategy1'][0]
str2 = df_match['Strategy2'][0]
df_match_s = df_match.loc[:, ["Round", "Period", "Action1", "Action2"]]
df_match_np = df_match_s.as_matrix()
periods = np.arange(period_max)
action1 = np.zeros((period_max, 2), dtype=int)
action2 = np.zeros((period_max, 2), dtype=int)
for r in range(round_len):
df_round = df_match_np[df_match_np[:, 0] == r]
period_len = int(df_round[:, 1].max()+1)
for p in range(period_len):
stage_df = df_round[df_round[:, 1] == p]
stage_action1 = stage_df[:, 2]
stage_action2 = stage_df[:, 3]
if stage_action1 == 0:
action1[p, 0] += 1
else:
action1[p, 1] += 1
if stage_action2 == 0:
action2[p, 0] += 1
else:
action2[p, 1] += 1
action1 = normalize(action1)
action2 = normalize(action2)
__plotC__(periods, action1[:, 0], action2[:, 0], period_max=100, title='{0} VS {1}'.format(str1, str2))
# 結果の読み込み
df_perfect = pd.read_csv("record_perfect.csv")
plotC(df_perfect)
In [9]:
def __plotAve__(ts_length, ave_payoff1, ave_payoff2, **kwargs):
title = kwargs.get('title', None)
fig = plt.figure(figsize=(18, 5))
weight1 = np.ones_like(ave_payoff1)/ts_length/100
ax1 = plt.subplot(1, 2, 1)
ax1.hist(ave_payoff1, bins=20, range=(0, 5), weights=weight1, normed=False, facecolor='g', alpha=0.8, label='player1')
plt.legend()
ax1.grid(True)
weight2 = np.ones_like(ave_payoff2)/ts_length/100
ax2 = plt.subplot(1, 2, 2, sharex=ax1, sharey=ax1)
ax2.hist(ave_payoff2, bins=20, range=(0, 5), weights=weight2, normed=False, facecolor='g', alpha=0.8, label='player2')
plt.legend()
ax2.grid(True)
mu1 = np.mean(ave_payoff1)
mu2 = np.mean(ave_payoff2)
ax1.axvline(x=mu1, linewidth=2, color='r')
ax2.axvline(x=mu2, linewidth=2, color='r')
plt.ylim(0, 1)
ax1.text(0.1, 0.9, "average = {0:.2f}".format(mu1), size=12)
ax2.text(0.1, 0.9, "average = {0:.2f}".format(mu2), size=12)
if title:
plt.suptitle(title, size=14)
plt.show()
def plotAve(df):
# 対戦組み合わせの数
match_len = df['Match'].max() + 1
# 1対戦のゲーム数
round_len = df['Round'].max() + 1
# 最長period
period_max = df['Period'].max() + 1
for m in range(match_len):
df_match = df[df.Match == m].reset_index()
str1 = df_match['Strategy1'][0]
str2 = df_match['Strategy2'][0]
df_match_s = df_match.loc[:, ["Round", "Period", "Payoff1", "Payoff2"]]
df_match_np = df_match_s.as_matrix()
periods = np.arange(period_max)
ave_payoff1 = np.zeros(round_len, dtype=float)
ave_payoff2 = np.zeros(round_len, dtype=float)
for r in range(round_len):
df_round = df_match_np[df_match_np[:, 0] == r]
period_len = int(df_round[:, 1].max()+1)
ave_payoff1[r] = df_round[:, 2].mean()
ave_payoff2[r] = df_round[:, 3].mean()
__plotAve__(ts_length, ave_payoff1, ave_payoff2, title='{0} / {1} average payoff distribution'.format(str1, str2))
# 結果の読み込み
df_perfect = pd.read_csv("record_perfect.csv")
plotAve(df_perfect)
In [10]:
# プロジェクトが成功か失敗かを返す
def public_signal(actions):
prob = rs.uniform()
if actions[0] == 0 and actions[1] == 0:
return 0 if prob < 0.9 else 1
elif (actions[0] == 0 and actions[1] == 1) or (actions[0] == 1 and actions[1] == 0):
return 0 if prob < 0.5 else 1
elif actions[0] == 1 and actions[1] == 1:
return 0 if prob < 0.2 else 1
else:
raise ValueError
strategies = [Iida_ipm, KatoStrategy, Self_Centered_public,
ImPubStrategy, GrimTrigger, MyStrategy, beeleb, OyamaImperfectPublicMonitoring, ogawa, yamagishi]
game = pl.RepeatedMatrixGame(payoff, strategies, signal=public_signal, ts_length=ts_length, repeat=1000)
game.play(mtype="public", random_seed=seed, record=False)
In [11]:
score_public = np.array([[0, 54398, 157243, 51002, 127446, 142462, 138347, 87568, 100961, 106747],
[95068, 0, 170665, 66666, 134419, 147247, 134836, 87358, 105889, 101710],
[57688, 4000, 0, 0, 117668, 136032, 137264, 70068, 81900, 115748],
[95162, 71666, 171665, 0, 134240, 147719, 135476, 86936, 106589, 100757],
[74731, 28164, 142248, 24950, 0, 137083, 138749, 75810, 90198, 114085],
[67987, 19612, 137657, 15964, 117708, 0,136851, 75450, 87852, 116362],
[69502, 27886, 137349, 24126, 117904, 136161, 0, 82602, 90045, 115409],
[97498, 59538, 154148, 56486, 129200, 141190, 134897, 0, 104634, 109458],
[87296, 47184, 151190, 43384, 125223, 140727, 137015, 92504, 0, 110338],
[89082, 49970, 142728, 47272, 127345, 135922, 135634, 89598, 96163, 0]], dtype=float)
score_public /= 34333
score_public = pd.DataFrame(score_public)
print(score_public)
| Iida_ipm | KatoStrategy | Self_Centered_public | ImPubStrategy | GrimTrigger | MyStrategy | beeleb | OyamaImperfectPublic | ogawa | yamagishi | |
|---|---|---|---|---|---|---|---|---|---|---|
| Iida_ipm | 0.000 | 1.584 | 4.580 | 1.486 | 3.712 | 4.149 | 4.030 | 2.551 | 2.941 | 3.109 |
| KatoStrategy | 2.769 | 0.000 | 4.971 | 1.942 | 3.915 | 4.289 | 3.927 | 2.544 | 3.084 | 2.962 |
| Self_Centered_public | 1.680 | 0.117 | 0.000 | 0.000 | 3.427 | 3.962 | 3.998 | 2.041 | 2.385 | 3.371 |
| ImPubStrategy | 2.772 | 2.087 | 5.000 | 0.000 | 3.910 | 4.303 | 3.946 | 2.532 | 3.105 | 2.935 |
| GrimTrigger | 2.177 | 0.820 | 4.143 | 0.727 | 0.000 | 3.993 | 4.041 | 2.208 | 2.627 | 3.323 |
| MyStrategy | 1.980 | 0.571 | 4.009 | 0.465 | 3.428 | 0.000 | 3.986 | 2.198 | 2.559 | 3.389 |
| beeleb | 2.024 | 0.812 | 4.000 | 0.703 | 3.434 | 3.966 | 0.000 | 2.406 | 2.623 | 3.361 |
| OyamaImperfectPublic | 2.840 | 1.734 | 4.490 | 1.645 | 3.763 | 4.112 | 3.929 | 0.000 | 3.048 | 3.188 |
| ogawa | 2.543 | 1.374 | 4.404 | 1.264 | 3.647 | 4.099 | 3.991 | 2.694 | 0.000 | 3.214 |
| yamagishi | 2.595 | 1.455 | 4.157 | 1.377 | 3.709 | 3.959 | 3.951 | 2.610 | 2.801 | 0.000 |
In [12]:
# 結果の読み込み
df_public = pd.read_csv("record_imperfect_public.csv")
plotC(df_public)
In [13]:
# 結果の読み込み
df_public = pd.read_csv("record_imperfect_public.csv")
plotAve(df_public)
In [3]:
# 「相手の」シグナルが協調か攻撃かを（ノイズ付きで）返す
def private_signal(actions):
pattern = [[0, 0], [0, 1], [1, 0], [1, 1]]
# 例えば実際の行動が(0, 1)なら、シグナルは(1, 0)である可能性が最も高い
signal_probs = [[.9, .02, .02, .06], [.02, .06, .9, .02], [.02, .9, .06, .02], [.06, .02, .02, .9]]
prob = rs.uniform()
if actions[0] == 0 and actions[1] == 0:
choice = rs.choice(4, p=signal_probs[0])
return pattern[choice]
elif actions[0] == 0 and actions[1] == 1:
choice = rs.choice(4, p=signal_probs[1])
return pattern[choice]
elif actions[0] == 1 and actions[1] == 0:
choice = rs.choice(4, p=signal_probs[2])
return pattern[choice]
elif actions[0] == 1 and actions[1] == 1:
choice = rs.choice(4, p=signal_probs[3])
return pattern[choice]
else:
raise ValueError
strategies = [Iida_iprm, KatoStrategy, Self_Centered_private,
ImPrivStrategy, GrimTrigger, MyStrategy, beeleb, OyamaImperfectPrivateMonitoring, ogawa, yamagishi]
game = pl.RepeatedMatrixGame(payoff, strategies, signal=private_signal, ts_length=ts_length, repeat=1000)
game.play(mtype="private", random_seed=seed, record=False)