def main():
from fireplace import cards
cards.db.initialize()
from utils import Agent,play_set_of_games,play_MechaHunterGames
from hearthstone.enums import CardClass
Human=Agent("Human",None,myClass=CardClass.MAGE)
StandardRandom=Agent("Maya",None) # Classを指定しないとHUNTER
# モンテカルロによる読み切り
## Maya=Agent("Maya",None)
# Standardなベクトル評価のエージェント34次元の重みベクトルをオプションとして入力する
#from agent_Standard import StandardStep1
#import random
#opt = []
#for i in range(34):
# opt.append(random.randint(1,10))
#GhostCatPlayer=Agent("GhostCat", StandardStep1,myOption=opt,myClass=CardClass.WARRIOR)
#あそび
#PIPlayer=Agent("Pi", StandardStep1,myOption=[3,1,4,1,5,9,2,6,5,3,5,8,9,7,9,3,2,3,8,4,6,2,6,4,3,3,8,3,2,7,9,5,0,2,8],myClass=CardClass.PRIEST)
# 言葉で戦略を組み立てるエージェント by Ahara
#from agent_word_strategy import WS, agent_word_strategy
#WSplayer = Agent("WS", agent_word_strategy,\
# myOption=[WS.ミニョンで敵ヒーローの体力を削る, WS.呪文を使えるなら呪文, WS.ランダムにプレー],\
# myClass=CardClass.PRIEST)
#AngryCat by Ahara 無駄な行為を選択肢から排除するアルゴリズム
#from agent_AngryCat import AngryCatAI
#AngryCatPlayer = Agent("AngryCat", AngryCatAI)
#ゲームプレイ
play_set_of_games(Human, StandardRandom, gameNumber=1, debugLog=True)
#ハンター縛りのデッキ（メカハンター）による対戦
#play_MechaHunterGames(StandardPlayer, AngryCatPlayer, gameNumber=1, debugLog=True)
#総当たり戦
#from competition import play_round_robin_competition
#play_round_robin_competition([StandardRandom,PIPlayer,AngryCatPlayer],matchNumber=5)
#特定の2枚のカードのシナジーを調べる(idea by Maya)
#from card_pair import investigate_card_pair, find_card_pair
#investigate_card_pair()
#シナジーのあるカードの組を漠然と探す
#find_card_pair(1)
#print("test_branch_yamadamaya")
pass
def getTransportistas(agentType, directoryAgent, sender, messageCount):
gmess = Graph()
# Construimos el mensaje de registro
gmess.bind('foaf', FOAF)
gmess.bind('dso', DSO)
ask_obj = agn[sender.name + '-Search']
gmess.add((ask_obj, RDF.type, DSO.Search))
gmess.add((ask_obj, DSO.AgentType, agentType))
gr = send_message(
build_message(gmess,
perf=ACL.request,
sender=sender.uri,
receiver=directoryAgent.uri,
msgcnt=messageCount,
content=ask_obj), directoryAgent.address)
dic = get_message_properties(gr)
content = dic['content']
agents = []
for (s, p, o) in gr.triples((content, None, None)):
if str(p).startswith('http://www.w3.org/1999/02/22-rdf-syntax-ns#_'):
address = gr.value(subject=o, predicate=DSO.Address)
url = gr.value(subject=o, predicate=DSO.Uri)
name = gr.value(subject=o, predicate=FOAF.name)
agent = Agent(name, url, address, None)
agents += [agent]
return agents
class MyForm(BoxLayout): # 此处类定义虽然为空，但会将my.kv的GUI定义的相关“程序”引入，即相当于在此定义
text_input = ObjectProperty() # 在类中添加text_input属性，对应kv文件中用于外部引用的名称，最终指向对应id的GUI部件
label_output = ObjectProperty() # 在类中添加label_output属性，对应kv文件中用于外部引用的名称，最终指向对应id的GUI部件
agent = Agent()
translate_res = None
# 加载字体资源（使用中文）
kivy.resources.resource_add_path("./fonts")
font_zh = kivy.resources.resource_find("msyh.ttc")
# 通过labelBase
LabelBase.register("msyh_labelBase", "msyh.ttc")
kivy.core.text.Label.register("msyh_label", "msyh.ttc")
def button_act(self, action=None):
print('Start ...')
if action is None:
self.translate_res = self.text_input.text # 获取text_input所指向GUI部件的text值，
elif action == 'Translate':
t = self.target_lang.text
if t == 'Target':
self.translate_res = self.text_input.text
else:
self.translate_res = self.agent.translate(text=self.text_input.text, source='auto', target=t)
elif action == 'Audio':
text = self.agent.transcribe()
t = self.target_lang.text
if t == 'Target':
t = 'en' # default
self.translate_res = self.agent.translate(text=text, source='auto', target=t)
elif action == 'Speech':
# 语音输出翻译结果
try:
assert self.translate_res is not None
self.agent.speech(self.translate_res)
print('Finish speeching ...')
except:
print('No speeching content.')
pass
else:
raise NotImplementedError
print(self.translate_res) # 打印结果到控制台
if self.translate_res is not None:
if action == 'Speech':
pass
else:
# 显示翻译结果到UI界面
self.label_output.text = self.translate_res
print('Finish translating ...')
pass
return
def clean_label(self,):
# 清除label文本
self.label_output.text = "" # (Waiting ...)
self.translate_res = None
return
def simulate_hunterschoice_policy_from_func(policy, N=1000):
grounds = create_small_hunting_environment()
env = HuntingMDPWrapper(grounds)
agent = Agent(env, policy)
successes = 0
episode_lengths = []
for i in range(N):
rewards = run_agent(agent)
if rewards[-1] == 1:
successes += 1
episode_lengths.append(len(rewards))
return successes / N, episode_lengths
def simulate_lake_policy_from_func(policy, N=1000, size="small"):
if size == "small":
env = get_small_lake()
else:
env = get_large_lake()
agent = Agent(env, policy)
successes = 0
episode_lengths = []
for i in range(N):
rewards = run_agent(agent)
if rewards[-1] == 1:
successes += 1
episode_lengths.append(len(rewards))
return successes / N, episode_lengths
def simulate_hunterschoice_policy(policy_location, N=1000):
print(f"Simulating Policy {policy_location}")
with open(policy_location, "rb") as f:
solver = pickle.load(f)
grounds = create_small_hunting_environment()
policy = create_hunterschoice_policy(solver.policy)
agent = Agent(grounds, policy)
successes = 0
episode_lengths = []
for i in range(N):
rewards = run_agent(agent)
if rewards[-1] == 1:
successes += 1
episode_lengths.append(len(rewards))
return successes / N, episode_lengths
def simulate_frozenlake_policy(policy_location, N=1000):
print(f"Simulating Policy {policy_location}")
with open(policy_location, "rb") as f:
solver = pickle.load(f)
if "small" in policy_location:
lake = get_small_lake()
else:
lake = get_large_lake()
policy = create_frozenlake_policy(solver.policy)
agent = Agent(lake, policy)
successes = 0
episode_lengths = []
for i in range(N):
rewards = run_agent(agent)
if rewards[-1] == 1:
successes += 1
episode_lengths.append(len(rewards))
return successes / N, episode_lengths
def getAgentInfo(agentType, directoryAgent, sender, messageCount):
gmess = Graph()
gmess.bind('foaf', FOAF)
gmess.bind('dso', DSO)
ask_obj = agn[sender.name + '-Search']
gmess.add((ask_obj, RDF.type, DSO.Search))
gmess.add((ask_obj, DSO.AgentType, agentType))
gr = send_message(
build_message(gmess,
perf=ACL.request,
sender=sender.uri,
receiver=directoryAgent.uri,
msgcnt=messageCount,
content=ask_obj), directoryAgent.address)
dic = get_message_properties(gr)
content = dic['content']
address = gr.value(subject=content, predicate=DSO.Address)
url = gr.value(subject=content, predicate=DSO.Uri)
name = gr.value(subject=content, predicate=FOAF.name)
return Agent(name, url, address, None)
RUN_ID = (
"cliff-GAMMA(%.2f)-ALPHA(%f)-MEMSIZE(%d)-EPISODES(%d)-BATCHSIZE(%d)-REPLACE_AFTER(%d)-PER(%r)"
% (GAMMA, ALPHA, MAX_MEMORY_SIZE, NUM_EPISODES, BATCH_SIZE, REPLACE_AFTER,
USE_PER))
time_started = datetime.now().strftime("%d.%m.%Y-%H:%M:%S")
if USE_TB: writer = SummaryWriter("runs/%s-%s" % (RUN_ID, time_started))
env = CliffWalking1D(9)
agent = Agent(Net=Net,
input_size=1,
output_size=env.action_space.n,
gamma=GAMMA,
alpha=ALPHA,
max_memory_size=MAX_MEMORY_SIZE,
num_episodes=NUM_EPISODES,
replace_after=REPLACE_AFTER,
use_per=USE_PER,
per_a=1)
# Filling memory
print("Filling replay memory")
state = env.reset()
if USE_TB:
writer.add_graph(agent.Q_loc, torch.FloatTensor([state]).to(device))
for t in range(200):
action = env.action_space.sample()
next_state, reward, done, info = env.step(action)
agent.store_transition(state, action, reward, next_state, done)
EPS_DECAY = 200 # tokyo uni value = 200
TARGET_UPDATE = 10 # for updating the target network
# for the experiments bash script:
record_file = "../exp/rl_results_" + args.data_name + ".csv"
# File for recording episode durations - original code:
# record_file = "../exp/rl_results.csv"
# Random Seed
for seed in range(1, 101): # original range (1,6)
random.seed(seed)
torch.manual_seed(0)
# Initialize Agent
agent = Agent()
# Get number of actions
# n_actions = len(agent.action_space)
n_actions = res_dict["num_words"]
print(f"n_actions: {n_actions}")
# Get size of state
state_size = agent.get_state().size(1)
# Initialize policy and target network
hidden_size = 32
policy_net = DQN(state_size, n_actions, hidden_size).to(device)
target_net = DQN(state_size, n_actions, hidden_size).to(device)
target_net.load_state_dict(policy_net.state_dict())
target_net.eval()
learning_rate = 1e-2 # tokyo uni value = 1e-2
def get_agent_any_type(type_opps, name, policy_type, env):
if type_opps == "zoo":
return load_agent(name, policy_type, "zoo_ant_policy_2", env, 1)
elif type_opps == "const":
trained_agent = constant_agent_sampler()
trained_agent.load(name)
return trained_agent
elif type_opps == "lstm":
policy = LSTMPolicy(scope="agent_new",
reuse=False,
ob_space=env.observation_space.spaces[0],
ac_space=env.action_space.spaces[0],
hiddens=[128, 128],
normalize=True)
def get_action(observation):
return policy.act(stochastic=True, observation=observation)[0]
trained_agent = Agent(get_action, policy.reset)
with open(name, "rb") as file:
values_from_save = pickle.load(file)
for key, value in values_from_save.items():
var = tf.get_default_graph().get_tensor_by_name(key)
sess.run(tf.assign(var, value))
return trained_agent
elif type_opps == "our_mlp":
#TODO DO ANYTHING BUT THIS. THIS IS VERY DIRTY AND SAD :(
def make_env(id):
# TODO: seed (not currently supported)
# TODO: VecNormalize? (typically good for MuJoCo)
# TODO: baselines logger?
# TODO: we're loading identical policy weights into different
# variables, this is to work-around design choice of Agent's
# having state stored inside of them.
sess = utils.make_session()
with sess.as_default():
multi_env = env
attacked_agent = constant_agent_sampler(act_dim=8,
magnitude=100)
single_env = Gymify(
MultiToSingle(CurryEnv(multi_env, attacked_agent)))
single_env.spec = gym.envs.registration.EnvSpec('Dummy-v0')
# TODO: upgrade Gym so don't have to do thi0s
single_env.observation_space.dtype = np.dtype(np.float32)
return single_env
# TODO: close session?
#TODO DO NOT EVEN READ THE ABOVE CODE :'(
denv = SubprocVecEnv([functools.partial(make_env, 0)])
model = ppo2.learn(network="mlp",
env=denv,
total_timesteps=1,
seed=0,
nminibatches=4,
log_interval=1,
save_interval=1,
load_path=name)
stateful_model = StatefulModel(denv, model)
trained_agent = utils.Agent(action_selector=stateful_model.get_action,
reseter=stateful_model.reset)
return trained_agent
raise (Exception('Agent type unrecognized'))