def get_act_afterstates(self, states):
tmp_game = Geister2()
max_num = self.bttl_num
mat = np.zeros((len(states), max_num))
for num in range(max_num):
for act_i in range(len(states)):
# tmp_gameに位置情報と自分の駒の色を設定
num_red = 0 # 敵の赤駒の数
is_vld = [] # 敵の盤上にある駒のi
for i in range(16):
tmp_game.units[i].x = self._game.units[i].x
tmp_game.units[i].y = self._game.units[i].y
tmp_game.units[i].taken = self._game.units[i].taken
tmp_game.units[i].color = self._game.units[i].color
if (i >= 8):
tmp_game.units[i].color = 1 # 敵の青駒として設定
if (tmp_game.units[i].taken is False):
is_vld.append(i)
# 敵の赤駒ならばnum_redを1追加
num_red += 1 if tmp_game.units[i].color == 3 else 0
# 敵の駒の色は推定せず，ランダムに
self._rnd.shuffle(is_vld)
for i in is_vld[:num_red]:
tmp_game.units[i].color = 3 # 敵の赤駒として設定
tmp_game.on_action_number_received(act_i)
mat[act_i, num] = battle_from(self.policy,
self.policy,
tmp_game=tmp_game)
means = mat.mean(axis=1)
return np.argmax(means)
def test():
seed = 2
game = Geister2()
tdagent = MCAgent(game, seed)
tdagent.w = np.array([
0.9, 0, 0, 0, 0, 0,
0.8, 0, 0, 0, 0, 0,
0.7, 0, 0, 0, 0, 0,
0.6, 0, 0, 0, 0, 0,
0.5, 0, 0, 0, 0, 0,
0.1, 0, 0, 0, 0, 0,
0, 0, 0, 0, 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, 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, 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, 0,
0, 0, 0, 0, 0, 0
])
rndagent = RandomAgent(game, seed)
agents = (tdagent, rndagent)
arr0, arr1 = (agent.init_red() for agent in agents)
game.setRed(arr0)
game.changeSide()
game.setRed(arr1)
game.changeSide()
game.printBoard()
player = 0
while not game.is_ended():
agent = agents[player]
states = game.after_states()
i_act = agent.get_act_afterstates(states)
game.on_action_number_received(i_act)
if player == 0:
game.printBoard()
game.changeSide()
player = (player+1) % 2
def battle2():
seed = 301
bttl_num = 50
game = Geister2()
agents = [[MCAgent(game, seed + i) for i in range(8)],
[MCAgent(game, seed + i + 8) for i in range(8)]]
agents_str = [
"weights_2/td_learned2_" + str(i) + ".npy" for i in range(1, 9)
]
for agent, string in zip(agents[0], agents_str):
agent.w = load(string)
agents_str = [
"weights_3/td_learned3_" + str(i) + ".npy" for i in range(1, 9)
]
for agent, string in zip(agents[1], agents_str):
agent.w = load(string)
means = np.zeros(8 * 8).reshape(8, 8)
for i in range(8):
for j in range(8):
r_list = np.zeros(bttl_num)
for t in range(bttl_num):
agent_s = (agents[0][i], agents[1][j])
arr0, arr1 = (agent.init_red() for agent in agent_s)
game.__init__()
game.setRed(arr0)
game.changeSide()
game.setRed(arr1)
game.changeSide()
player = 0
while not game.is_ended():
agent = agent_s[player]
states = game.after_states()
i_act = agent.get_act_afterstates(states)
game.on_action_number_received(i_act)
game.changeSide()
player = (player + 1) % 2
if player == 1:
game.changeSide()
result = game.checkResult()
r = (1 if (result > 0) else (-1 if (result < 0) else 0))
r_list[t] = r
means[i][j] = r_list.mean()
print(means)
print("mean: ", means.mean())
def learn():
file_name = "td_9"
seed = 91
game = Geister2()
mcagent = MCAgent(game, seed)
opponent = RandomAgent(game, seed+1)
env = VsEnv(opponent, game, seed)
mcagent.learn(env, seed)
# for k in range(6*7*3):
# for i in range(3):
# for j in range(7):
# print((mcagent.w[j+i*(6*7)+k*(6*7*3):6+j+i*(6*7)+k*(6*7*3)]
# * 1000).round()*(1/1000))
# print("-----------")
# print("-------------------")
np.save(file_name, mcagent.w)
w_td = np.load(file_name+'.npy')
print(w_td.shape)
def battle():
seed = 29
bttl_num = 10
game = Geister2()
agents_str = [
"weights/weights_13/reinforce_" + str(i) + "_theta.npy"
for i in range(1, 9)
]
agent_len = len(agents_str)
agents = [REINFORCEAgent(game, seed + i) for i in range(agent_len)]
for agent, string in zip(agents, agents_str):
agent.theta = load(string)
means = np.zeros((agent_len, agent_len))
for i in range(len(agents)):
for j in range(i, len(agents)):
if i == j:
continue
r_list = np.zeros(bttl_num)
for t in range(bttl_num):
agent_s = (agents[i], agents[j])
arr0, arr1 = (agent.init_red() for agent in agent_s)
game.__init__()
game.setRed(arr0)
game.changeSide()
game.setRed(arr1)
game.changeSide()
player = 0
while not game.is_ended():
agent = agent_s[player]
states = game.after_states()
i_act = agent.get_act_afterstates(states)
game.on_action_number_received(i_act)
game.changeSide()
player = (player + 1) % 2
if player == 1:
game.changeSide()
result = game.checkResult()
r = (1 if (result > 0) else (-1 if (result < 0) else 0))
r_list[t] = r
means[i][j] = r_list.mean()
means[j][i] = -means[i][j]
print(means)
def learn():
file_name = "weights/rfvsrnd6"
seed = 103
game = Geister2()
agent = REINFORCEAgent(game, seed)
agent.w = np.random.randn(agent.W_SIZE) * agent.alpha * 0.0001
agent.theta = np.random.randn(agent.T_SIZE) * agent.beta * 0.0001
opponent = RandomAgent(game, seed + 1)
env = VsEnv(opponent, game, seed)
# 計測準備
pr = cProfile.Profile()
pr.enable()
# 計測開始
agent.learn(env, seed)
# 計測終了，計測結果出力
pr.disable()
stats = pstats.Stats(pr)
stats.sort_stats('cumtime')
stats.print_stats()
pr.dump_stats('profile.stats')
# 事後処理
np.save(file_name + "_w", agent.w)
np.save(file_name + "_theta", agent.theta)
def test2():
mcagent = MCAgent(Geister2())
mcagent.w = np.load("td_4.npy")
print(mcagent.init_red())
def ranking_learn(game):
# ranking_leanのデータ読み込み
with open(ranking_data_path, 'rt') as fin:
cin = csv.reader(fin)
datas = [row for row in cin if len(row) > 0]
num_weights = int(datas[0][0])
# ランキングデータの読み込み
ranking_path = []
ranking_n = []
ranking_r = []
with open(rankings_path, 'rt') as fin:
cin = csv.reader(fin)
datas = [row for row in cin if len(row) > 0]
ranking_path = [row[2] for row in datas]
ranking_n = [int(row[3]) for row in datas]
ranking_r = [float(row[4]) for row in datas]
# ランキングの人数が足りないときはランダムな重みのREINFORCEagentを追加
while len(ranking_path) < num_rankingagents:
ranking_path.append(get_path_radom(game)) # path_listの末尾を移動
ranking_n.append(0)
ranking_r.append(0)
# ランキングの人数が多すぎるときはエラー
if len(ranking_path) > num_rankingagents:
print("error. ranking_num is over num_rankingagents")
# path_list.append(ranking_path.pop(0)) # rankingの末尾を移動
# del ranking_n[0]
# del ranking_r[0]
# rank agentsの重みの読み込み
game = Geister2()
train_is = rnd.sample(range(num_rankingagents), num_rankingagents // 2)
test_is = [i for i in range(num_rankingagents) if i not in train_is]
rank_agents = load_agents(ranking_path, game, None)
train_agents = [rank_agents[i] for i in train_is]
test_agents = [rank_agents[i] for i in test_is]
# 新しいagentの作成
agent = pick_agent(game)
agent_path = weights_path + "/rankRF" + str(num_weights)
# agnetの学習
env = VsEnvs(train_agents, game, None) # 対戦相手はランダムに一度だけ
agent.learn(env, max_episodes=max_episodes)
# 最新のランキングに対して改めて対戦を行い，(test_agentsのみ更新)
# 基準を満たしていれば，agentのランキングへの追加
results = []
for i in test_is:
test_agent = rank_agents[i]
# resultはagentの勝率
result = battle(agent, test_agent, bttl_num=bttl_num, seed=None)
results.append(result)
# 対戦相手の勝率を更新
r_opp = -result
ranking_r[i] = (ranking_r[i] * ranking_n[i] + r_opp) / (ranking_n[i] +
1)
ranking_n[i] += 1
results = np.array(results)
# 基準を満たしている場合(rが一定値以上かつ過半数に対し勝利),ランキングに追加
if (results.mean() > threshold
and len(np.where(results > 0)[0]) > num_rankingagents / 2):
# ランキングの削除対象(test_agentsのうち勝率が最低のもの)
dl_index = ranking_r.index(min([ranking_r[i] for i in test_is]))
ranking_path[dl_index] = agent_path
ranking_n[dl_index] = ranking_r[dl_index] = 0
# agentのデータの書き込み
np.save(agent_path + "_w", agent.w)
np.save(agent_path + "_theta", agent.theta)
num_weights += 1
# ranking_learnのデータ書き込み
with open(ranking_data_path, 'wt') as fout:
csvout = csv.writer(fout)
datas = [[str(num_weights)]]
csvout.writerows(datas)
# ランキングデータの書き込み
datas = [[str(i + 1), "REINFORCEAgent", ranking_path[i], n, r]
for i, n, r in zip(range(len(ranking_path)), ranking_n, ranking_r)
]
with open(rankings_path, 'wt') as fout:
csvout = csv.writer(fout)
csvout.writerows(datas)
# ranking_learnのデータ書き込み
with open(ranking_data_path, 'wt') as fout:
csvout = csv.writer(fout)
datas = [[str(num_weights)]]
csvout.writerows(datas)
# ランキングデータの書き込み
datas = [[str(i + 1), "REINFORCEAgent", ranking_path[i], n, r]
for i, n, r in zip(range(len(ranking_path)), ranking_n, ranking_r)
]
with open(rankings_path, 'wt') as fout:
csvout = csv.writer(fout)
csvout.writerows(datas)
if __name__ == "__main__":
game = Geister2()
# 計測準備
pr = cProfile.Profile()
pr.enable()
# 計測開始
while (True):
ranking_learn(game)
# 計測終了，計測結果出力
pr.disable()
stats = pstats.Stats(pr)
stats.sort_stats('cumtime')
stats.print_stats()
pr.dump_stats('profile.stats')
# 結果出力終了
def cluster_learn():
seed = 122
file_name = "weights/weights_16/reinforce_"
agents_len = 18
max_episodes = 500 * (agents_len)
plt_intvl = 50 * (agents_len)
plt_bttl = 200
linestyles = [':', '--', '-.'] # alphaに相当 # linestyle=(0, (1, 0))
plt_colors = ['m', 'r', 'g', 'c', 'b', 'y'] # betaに相当,mマゼンタ(紫),cシアン(青緑)
linestyle_avg = '-'
plt_color_avg = 'k'
alphas = [0.003, 0.005, 0.01]
betas = [0.0005, 0.0001, 0.0003, 0.0005, 0.001, 0.0015]
assert (len(linestyles) == len(alphas))
assert (len(plt_colors) == len(betas))
assert (len(alphas) * len(betas) == agents_len)
game = Geister2()
np.random.seed(seed)
rnd = random.Random(seed)
agents = [REINFORCEAgent(game, seed + i) for i in range(agents_len)]
for i in range(len(alphas)):
for j in range(len(betas)):
agents[i + j * len(alphas)].alpha = alphas[i]
agents[i + j * len(alphas)].beta = betas[j]
# 重みを小さな正規乱数で初期化
for agent in agents:
if agent.w is None:
agent.w = np.zeros(agent.W_SIZE)
if agent.theta is None:
agent.theta = np.zeros(agent.T_SIZE)
episodes_x = []
results_y = [[] for _ in range(agents_len)]
avg_y = []
rnd_agent = RandomAgent(game, seed * 2 + 1)
env = VsEnv(agents[0], game, seed)
for episode in range(max_episodes):
# 学習個体を一度ずつ選ぶ(順番はランダム)
for i in rnd.sample(range(agents_len), agents_len): # -> [2, 0, 1]など
# i = rnd.randrange(agents_len) # 学習個体はランダム
# # 対戦相手は，全ての候補を一度ずつ選ぶ(順番はランダム)
# for j in rnd.sample(range(agents_len), agents_len):
j = rnd.randrange(agents_len) # 対戦相手はランダムに一度だけ
agent = agents[i]
env._opponent = agents[j]
agent.learn(env, max_episodes=1)
# 定期的にランダムとの対戦結果を描画
if (episode + 1) % plt_intvl == 0:
episodes_x.append(episode)
plt.clf()
opponent = rnd_agent
env._opponent = opponent
avgs = []
for i in range(agents_len):
agent = agents[i]
theta = agent.theta
r_list = np.zeros(plt_bttl)
for bttl_i in range(plt_bttl):
afterstates = env.on_episode_begin(agent.init_red())
x = agent.get_x(afterstates)
a = agent.get_act(x, theta)
for t in range(300):
r, nafterstates = env.on_action_number_received(a)
if r != 0:
break
nx = agent.get_x(nafterstates)
na = agent.get_act(nx, theta)
x = nx
a = na
r_list[bttl_i] = r
mean = r_list.mean()
avgs.append(mean)
results_y[i].append(mean)
plt.figure(1)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Results')
x_list = np.array(episodes_x)
y_list = np.array(results_y[i])
plt.plot(x_list,
y_list,
linestyle=linestyles[i % len(alphas)],
c=plt_colors[i // len(alphas)],
label=str(i))
avg_y.append(np.array(avgs).mean())
plt.figure(1)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Results')
x_list = np.array(episodes_x)
y_list = np.array(avg_y)
plt.plot(x_list,
y_list,
linestyle=linestyle_avg,
c=plt_color_avg,
label=agents_len)
plt.pause(0.01) # pause a bit so that plots are updated
plt.savefig(file_name + str(".png"))
plt.show()
for i in range(agents_len):
np.save(file_name + str(i + 1) + "_w", agents[i].w)
np.save(file_name + str(i + 1) + "_theta", agents[i].theta)
def setUp(self):
game = Geister2()
game.setRed(["E", "F", "G", "H"])
game.changeSide()
game.setRed(["E", "F", "G", "H"])
self.game = game
def learn(self,
env,
seed=1,
max_episodes=100000,
draw_mode=False,
draw_opp=None):
alpha = self.alpha
beta = self.beta
# epsilon = self.epsilon
# rnd = self._rnd
assert (env.S_SIZE == self.S_SIZE)
plt_intvl = 500
plt_bttl = 50
episodes_x = []
results_y = []
dlts_y = []
dlts = []
# 読み込み
# mcagent.w = np.load("td_4.npy")
# wを小さな正規乱数で初期化
# np.random.seed(seed)
# if self.w is None:
# self.w = np.random.randn(self.W_SIZE)*alpha*0.1
# if self.theta is None:
# self.theta = np.random.randn(self.T_SIZE)*beta*0.1
w = self.w
theta = self.theta
if draw_mode:
denv = VsEnv(draw_opp, game=Geister2(), seed=seed)
for episode in range(max_episodes):
afterstates = env.on_episode_begin(self.init_red())
xs = self.get_x([env.get_state()])[0]
x = self.get_x(afterstates)
a = self.get_act(x, theta)
xs_list = [xs]
x_list = [x]
xa_list = [x[a]]
for t in range(self.MAX_T):
r, nafterstates = env.on_action_number_received(a)
if r != 0:
break
nxs = self.get_x([env.get_state()])[0]
nx = self.get_x(nafterstates)
na = self.get_act(nx, theta)
xs_list.append(nxs)
x_list.append(nx)
xa_list.append(nx[na])
x = nx
a = na
for xa, x, xs in zip(xa_list, x_list, xs_list):
q = 2 / (1 + np.exp(-np.dot(w, xs))) - 1
dlt = r - q # 報酬予測は事後状態を用いてはならない
dlts.append(dlt**2)
w += beta * dlt * xs
hs = x.dot(theta)
hs -= hs.max() # overflow回避のため
exps = np.exp(hs)
pis = exps / exps.sum()
theta += alpha * r * (xa - pis.dot(x))
# 焼きなまし法
# theta += alpha*(episode/max_episodes)*r*(xa - pis.dot(x))
if draw_opp is None and draw_mode:
print("not implemented")
raise Exception
if draw_mode and ((episode + 1) % plt_intvl == 0):
dlts_y.append(np.array(dlts).mean())
dlts = []
if draw_opp is not None:
denv._opponent = draw_opp
r_sum = 0.0
for bttl_i in range(plt_bttl):
afterstates = denv.on_episode_begin(self.init_red())
x = self.get_x(afterstates)
a = self.get_act(x, theta)
for t in range(300):
r, nafterstates = denv.on_action_number_received(a)
if r != 0:
break
nx = self.get_x(nafterstates)
na = self.get_act(nx, theta)
x = nx
a = na
r_sum += r
results_y.append(r_sum / plt_bttl)
episodes_x.append(episode)
# 一つ目 results
plt.figure(2)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Results of Interval')
plt.text(50, 0.5, "alpha=" + str(self.alpha))
plt.text(50, 0.4, "beta=" + str(self.beta))
x_list = np.array(episodes_x)
y_list = np.array(results_y)
plt.plot(x_list, y_list)
plt.pause(0.0001) # pause a bit so that plots are updated
plt.clf()
# 二つ目 予測誤差 Δv(s)^2
plt.figure(1)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Dlt v(s)^2')
plt.text(50, 0.5, "alpha=" + str(self.alpha))
plt.text(50, 0.4, "beta=" + str(self.beta))
x_list = np.array(episodes_x)
y_list = np.array(dlts_y)
plt.plot(x_list, y_list)
plt.pause(0.0001) # pause a bit so that plots are updated
plt.clf()
# 学習終了後
if (draw_mode):
# 一つ目 results
plt.figure(2)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Results of Interval')
plt.text(50, 0.5, "alpha=" + str(self.alpha))
plt.text(50, 0.4, "beta=" + str(self.beta))
x_list = np.array(episodes_x)
y_list = np.array(results_y)
plt.plot(x_list, y_list)
plt.show()
# 二つ目 予測誤差 Δv(s)^2
plt.figure(1)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Dlt v(s)^2')
plt.text(50, 0.5, "alpha=" + str(self.alpha))
plt.text(50, 0.4, "beta=" + str(self.beta))
x_list = np.array(episodes_x)
y_list = np.array(dlts_y)
plt.plot(x_list, y_list)
plt.show()
game.changeSide()
game.setRed(arr1)
game.changeSide()
player = 0
while not game.is_ended():
agent = agents[player]
states = game.after_states()
i_act = agent.get_act_afterstates(states)
game.on_action_number_received(i_act)
game.changeSide()
player = (player + 1) % 2
if player == 1:
game.changeSide()
result = game.checkResult()
r = (1 if (result > 0) else (-1 if (result < 0) else 0))
results[t] = r
return results.mean()
if __name__ == "__main__":
seed = 100
geister = Geister2()
agents1 = [load_agent("weights/rfvsrnd5", geister, seed)]
agents2 = [
load_agent(("weights/weights_13/reinforce_" + str(i)), geister, seed)
for i in range(1, 10)
]
results = battle2(agents1, agents2, 100)
print(results)
def __init__(self, opponent, game=Geister2(), seed=0):
self._opponent = opponent
self._game = game
self._seed = seed
self.S_SIZE = (6 * 6 + 6) * 3
def self_play(file_name, agent=None, max_train=50000):
seed = 0
max_episodes = max_train
plt_intvl = max_episodes + 1 # プロットしない
plt_bttl = 200
linestyle = '-' # alphaに相当 # linestyle=(0, (1, 0))
plt_color = 'k' # betaに相当, # mマゼンタ(紫),cシアン(青緑)なども
alpha = 0.001
beta = 0.0001
game = Geister2()
np.random.seed(seed)
if agent is None:
agent = REINFORCEAgent(game, seed)
agent.w = np.zeros(agent.W_SIZE)
agent.theta = np.zeros(agent.T_SIZE)
agent.alpha = alpha
agent.beta = beta
episodes_x = []
results_y = []
rnd_agent = RandomAgent(game, seed*2+1)
env = VsEnv(agent, game, seed)
denv = VsEnv(rnd_agent, game, seed)
for episode in range(max_episodes):
agent.alpha = alpha # * (1 - episode/max_episodes)
agent.beta = beta # * (1 - episode/max_episodes)
agent.learn(env, max_episodes=1)
# 定期的にランダムとの対戦結果を描画
if (episode) % plt_intvl == 0:
episodes_x.append(episode)
plt.clf()
opponent = rnd_agent
denv._opponent = opponent
theta = agent.theta
r_list = np.zeros(plt_bttl)
for bttl_i in range(plt_bttl):
afterstates = denv.on_episode_begin(agent.init_red())
x = agent.get_x(afterstates)
a = agent.get_act(x, theta)
for t in range(300):
r, nafterstates = denv.on_action_number_received(a)
if r != 0:
break
nx = agent.get_x(nafterstates)
na = agent.get_act(nx, theta)
x = nx
a = na
r_list[bttl_i] = r
mean = r_list.mean()
results_y.append(mean)
plt.figure(1)
plt.title('Training...')
plt.xlabel('Episode')
plt.ylabel('Mean Results')
x_list = np.array(episodes_x)
y_list = np.array(results_y)
plt.plot(x_list, y_list,
linestyle=linestyle,
c=plt_color)
plt.pause(0.01) # pause a bit so that plots are updated
plt.savefig(file_name+str(".png"))
# self play向けに変更
np.save(file_name+"_w", agent.w)
np.save(file_name+"_theta", agent.theta)
# numpyに変換し，グラフの情報を保存
np.save(file_name+"x_list", np.array(x_list))
np.save(file_name+"results_y", np.array(results_y))
def __init__(self, opponents, game=Geister2(), seed=0):
self._opponents = opponents
opp = rnd.choice(opponents)
return super().__init__(opponent=opp, game=game, seed=seed)
def add_in_ranking(path_list):
# ランキングデータの読み込み
ranking_path = []
ranking_n = []
ranking_r = []
with open('rankings', 'rt') as fin:
cin = csv.reader(fin)
datas = [row for row in cin if len(row) > 0]
ranking_path = [row[2] for row in datas]
ranking_n = [int(row[3]) for row in datas]
ranking_r = [float(row[4]) for row in datas]
# ランキングの人数が足りないときはagentを追加
while len(ranking_path) < agentsnum:
ranking_path.append(path_list.pop(0)) # path_listの末尾を移動
ranking_n.append(0)
ranking_r.append(0)
# ランキングの人数が多すぎるときは追加候補に移動
while len(ranking_path) > agentsnum:
path_list.append(ranking_path.pop(0)) # rankingの末尾を移動
del ranking_n[0]
del ranking_r[0]
# agentの重みの読み込み
game = Geister2()
agents = load_agents(path_list, game, seed)
rank_agents = load_agents(ranking_path, game, seed)
# 対戦を行い基準を満たせば追加．
for i in range(len(agents)):
agent = agents[i]
results = []
# 最新のランキングに更新
rank_agents = load_agents(ranking_path, game, seed)
for j in range(len(rank_agents)):
rank_agent = rank_agents[j]
# resultはagentの勝率
result = battle(agent, rank_agent, bttl_num=bttl_num, seed=seed)
results.append(result)
# 対戦相手の勝率を更新
r_opp = -result
ranking_r[j] = (ranking_r[j] * ranking_n[j] +
r_opp) / (ranking_n[j] + 1)
ranking_n[j] += 1
results = np.array(results)
# 基準を満たしていない場合
if (results.mean() <= threshold
or len(np.where(results > 0)[0]) <= agentsnum / 2):
continue
# 基準を満たしている場合(rが一定値以上かつ過半数に対し勝利),ランキングに追加
dl_index = ranking_r.index(min(ranking_r)) # ランキングの削除対象
ranking_path[dl_index] = path_list[i]
ranking_n[dl_index] = ranking_r[dl_index] = 0
# データの書き込み
datas = [[str(i + 1), "REINFORCEAgent", ranking_path[i], n, r]
for i, n, r in zip(range(len(ranking_path)), ranking_n, ranking_r)
]
with open('rankings', 'wt') as fout:
csvout = csv.writer(fout)
csvout.writerows(datas)
