def __init__(self, path="data/arrange_model", messages=None, mode=1):
# some parameter
map_size = 250
eps = 0.15
# init the game
env = magent.GridWorld(load_config(map_size))
font = FontProvider('data/font_8x8/basic.txt')
handles = env.get_handles()
food_handle, handles = handles[0], handles[1:]
models = []
models.append(DeepQNetwork(env, handles[0], 'arrange', use_conv=True))
# load model
models[0].load(path, 10)
# init environment
env.reset()
generate_map(mode, env, map_size, food_handle, handles, messages, font)
# save to member variable
self.env = env
self.food_handle = food_handle
self.handles = handles
self.eps = eps
self.models = models
self.done = False
self.map_size = map_size
self.new_rule_ct = 0
self.pos_reward_ct = set()
self.num = None
self.ct = 0
def __init__(self,
path="data/battle_model",
total_step=1000,
add_counter=10,
add_interval=50):
# some parameter
map_size = 125
eps = 0.05
# init the game
env = magent.GridWorld(load_config(map_size))
handles = env.get_handles()
models = []
models.append(
DeepQNetwork(env,
handles[0],
'trusty-battle-game-l',
use_conv=True))
models.append(
DeepQNetwork(env,
handles[1],
'trusty-battle-game-r',
use_conv=True))
# load model
models[0].load(path, 0, 'trusty-battle-game-l')
models[1].load(path, 0, 'trusty-battle-game-r')
# init environment
env.reset()
generate_map(env, map_size, handles)
# save to member variable
self.env = env
self.handles = handles
self.eps = eps
self.models = models
self.map_size = map_size
self.total_step = total_step
self.add_interval = add_interval
self.add_counter = add_counter
self.done = False
print(env.get_view2attack(handles[0]))
plt.show()
# init the game "pursuit" (config file are stored in python/magent/builtin/config/)
env = magent.GridWorld("pursuit", map_size=map_size)
#
env.set_render_dir("build/render")
# get group handles
predator, prey = env.get_handles()
# init env and agents
env.reset()
env.add_walls(method="random", n=map_size * map_size * 0.01)
env.add_agents(predator, method="random", n=map_size * map_size * 0.02)
env.add_agents(prey, method="random", n=map_size * map_size * 0.02)
# init two models
model1 = DeepQNetwork(env, predator, "predator")
model2 = DeepQNetwork(env, prey, "prey")
# load trained model
model1.load("data/pursuit_model")
model2.load("data/pursuit_model")
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(predator), env.get_num(prey)))
while not done:
# take actions for deers
obs_1 = env.get_observation(predator)
ids_1 = env.get_agent_id(predator)
acts_1 = model1.infer_action(obs_1, ids_1)
env.set_action(predator, acts_1)
# init models
batch_size = 512
unroll_step = 8
target_update = 1200
train_freq = 5
models = []
if args.alg == 'dqn':
from magent.builtin.tf_model import DeepQNetwork
models.append(
DeepQNetwork(env,
handles[0],
"battle",
batch_size=batch_size,
learning_rate=3e-4,
memory_size=2**21,
target_update=target_update,
train_freq=train_freq,
eval_obs=eval_obs))
elif args.alg == 'drqn':
from magent.builtin.tf_model import DeepRecurrentQNetwork
models.append(
DeepRecurrentQNetwork(env,
handles[0],
"battle",
learning_rate=3e-4,
batch_size=batch_size / unroll_step,
unroll_step=unroll_step,
memory_size=2 * 8 * 625,
target_update=target_update,
deer_handle, tiger_handle = env.get_handles()
# init two models
models = [
RandomActor(env, deer_handle, tiger_handle),
]
batch_size = 512
unroll = 8
if args.alg == 'dqn':
from magent.builtin.tf_model import DeepQNetwork
models.append(
DeepQNetwork(env,
tiger_handle,
args.name,
batch_size=batch_size,
memory_size=2**20,
learning_rate=4e-4))
step_batch_size = None
elif args.alg == 'drqn':
from magent.builtin.tf_model import DeepRecurrentQNetwork
models.append(
DeepRecurrentQNetwork(env,
tiger_handle,
"tiger",
batch_size=batch_size / unroll,
unroll_step=unroll,
memory_size=20000,
learning_rate=4e-4))
step_batch_size = None
elif args.alg == 'a2c':
deer_handle, tiger_handle = env.get_handles()
env.add_walls(method="random", n=agent_number / 10)
env.add_agents(deer_handle, method="random", n=agent_number / 2)
env.add_agents(tiger_handle, method="random", n=agent_number / 2)
# init two models
if args.num_gpu == 0:
model1 = RandomActor(env, deer_handle, "deer")
model2 = RandomActor(env, tiger_handle, "tiger")
else:
if args.frame == 'tf':
from magent.builtin.tf_model import DeepQNetwork
else:
from magent.builtin.mx_model import DeepQNetwork
model1 = DeepQNetwork(env, deer_handle, "deer", num_gpu=args.num_gpu, infer_batch_size=100000)
model2 = DeepQNetwork(env, tiger_handle, "tiger", num_gpu=args.num_gpu, infer_batch_size=100000)
total_reward = 0
print(env.get_view_space(deer_handle))
print(env.get_view_space(tiger_handle))
total_time = 0
for i in range(n_step):
print("===== step %d =====" % i)
start_time = time.time()
obs_1 = measure_time("get obs 1", env.get_observation, deer_handle)
acts_1 = measure_time("infer act 1", model1.infer_action, obs_1, None)
relation1_t_ = merge(relation1_t, mode='concat', concat_axis=1)
relation2_t = []
for j in range(n_agent):
T = Lambda(lambda x: K.batch_dot(x[0], x[1]))(
[In_t[j * 2 + 1], relation1_t_])
relation2_t.append(m2_t([T, T, T, In_t[40]]))
V_t = []
for j in range(n_agent):
V_t.append(q_net_t([feature_t[j], relation1_t[j], relation2_t[j]]))
model_t = Model(input=In_t, output=V_t)
tf_model = DeepQNetwork(env, handles[1], 'trusty-battle-game-l', use_conv=True)
tf_model.load("data/battle_model", 0, 'trusty-battle-game-l')
###########playing#############
while i_episode < n_episode:
alpha *= 0.996
if alpha < 0.01:
alpha = 0.01
print(i_episode)
i_episode = i_episode + 1
env.reset()
#env.add_walls(method="random", n=map_size * map_size * 0.03)
env.add_agents(handles[0], method="random", n=20)
env.add_agents(handles[1], method="random", n=12)
step_ct = 0
done = False
n = len(handles)
"""rename tensorflow models"""
import sys
import magent
from magent.builtin.tf_model import DeepQNetwork
env = magent.GridWorld("battle", map_size=125)
handles = env.get_handles()
rounds = eval(sys.argv[1])
for i in [rounds]:
model = DeepQNetwork(env, handles[0], "battle")
print("load %d" % i)
model.load("data/", i, "selfplay")
print("save %d" % i)
model.save("data/battle_model", i)
# init the game "pursuit" (config file are stored in python/magent/builtin/config/)
env = magent.GridWorld("pursuit", map_size=map_size)
env.set_render_dir("build/render")
# get group handles
predator, prey = env.get_handles()
# init env and agents
env.reset()
env.add_walls(method="random", n=map_size * map_size * 0.01)
env.add_agents(predator, method="random", n=map_size * map_size * 0.02)
env.add_agents(prey, method="random", n=map_size * map_size * 0.02)
# init two models
model1 = DeepQNetwork(env, predator, "predator")
model2 = DeepQNetwork(env, prey, "prey")
# load trained model
model1.load("data/pursuit_model")
model2.load("data/pursuit_model")
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(predator), env.get_num(prey)))
while not done:
# take actions for deers
obs_1 = env.get_observation(predator)
ids_1 = env.get_agent_id(predator)
acts_1 = model1.infer_action(obs_1, ids_1)
env.set_action(predator, acts_1)
# init the game "pursuit" (config file are stored in python/magent/builtin/config/)
env = magent.GridWorld("battle", map_size=map_size)
env.set_render_dir("build/render")
# get group handles
army1, army2 = env.get_handles()
# init env and agents
env.reset()
env.add_walls(method="random", n=map_size * map_size * 0.02)
env.add_agents(army1, method="random", n=map_size * map_size * 0.05)
env.add_agents(army2, method="random", n=map_size * map_size * 0.05)
# init two models
model1 = DeepQNetwork(env, army1, "battle-l", memory_size=2 ** 10)
model2 = DeepQNetwork(env, army2, "battle-r", memory_size=2 ** 10)
# load trained model
model1.load("save_model", 9)
model2.load("save_model", 10)
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(army1), env.get_num(army2)))
result = []
while not done:
# take actions for deers
obs_1 = env.get_observation(army1)
ids_1 = env.get_agent_id(army1)
acts_1 = model1.infer_action(obs_1, ids_1)
env.reset()
env.add_walls(method="random", n=map_size * map_size * 0.02)
env.add_agents(army1, method="random", n=map_size * map_size * 0.05)
env.add_agents(army2, method="random", n=map_size * map_size * 0.05)
env.add_agents(army3, method="random", n=map_size * map_size * 0.05)
env.add_agents(army4, method="random", n=map_size * map_size * 0.05)
env.add_agents(army5, method="random", n=map_size * map_size * 0.05)
env.add_agents(army6, method="random", n=map_size * map_size * 0.05)
env.add_agents(army7, method="random", n=map_size * map_size * 0.05)
env.add_agents(army8, method="random", n=map_size * map_size * 0.05)
env.add_agents(army9, method="random", n=map_size * map_size * 0.05)
env.add_agents(army10, method="random", n=map_size * map_size * 0.05)
# init two models
# tf.reset_default_graph()
model1 = DeepQNetwork(env, army1, "battle-l", memory_size=2**10)
# tf.reset_default_graph()
model2 = DeepQNetwork(env, army2, "battle-r", memory_size=2**10)
model1.load("save_model", 18)
model2.load("save_model", 18)
# tf.reset_default_graph()
model3 = DeepQNetwork(env, army3, "battle-l", memory_size=2**10)
# tf.reset_default_graph()
model4 = DeepQNetwork(env, army4, "battle-r", memory_size=2**10)
model3.load("save_model", 17)
model4.load("save_model", 17)
# tf.reset_default_graph()
model5 = DeepQNetwork(env, army5, "battle-l", memory_size=2**10)
# tf.reset_default_graph()
eval_obs = magent.utility.sample_observation(env, handles, 2048,
500)[0]
# init models
batch_size = 256
unroll_step = 8
target_update = 1000
train_freq = 5
models = []
if args.alg == 'dqn':
models.append(
DeepQNetwork(env,
handles[0],
"selfplay",
batch_size=batch_size,
memory_size=2**20,
target_update=target_update,
train_freq=train_freq,
eval_obs=eval_obs))
elif args.alg == 'drqn':
models.append(
DeepRecurrentQNetwork(env,
handles[0],
"selfplay",
batch_size=batch_size / unroll_step,
unroll_step=unroll_step,
memory_size=2 * 8 * 625,
target_update=target_update,
train_freq=train_freq,
eval_obs=eval_obs))
else:
