def test_train(self):
memory_init_size = 20
num_steps = 1000
agent = NFSPAgent(num_actions=2,
state_shape=[2],
hidden_layers_sizes=[10, 10],
reservoir_buffer_capacity=50,
batch_size=4,
min_buffer_size_to_learn=memory_init_size,
q_replay_memory_size=50,
q_replay_memory_init_size=memory_init_size,
q_batch_size=4,
q_mlp_layers=[10, 10],
device=torch.device('cpu'))
predicted_action, _ = agent.eval_step({
'obs':
np.random.random_sample((2, )),
'legal_actions': {
0: None,
1: None
},
'raw_legal_actions': ['call', 'raise']
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
for _ in range(num_steps):
agent.sample_episode_policy()
predicted_action = agent.step({
'obs': np.random.random_sample((2, )),
'legal_actions': {
0: None,
1: None
}
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
ts = [{
'obs': np.random.random_sample((2, )),
'legal_actions': {
0: None,
1: None
}
},
np.random.randint(2), 0, {
'obs': np.random.random_sample((2, )),
'legal_actions': {
0: None,
1: None
},
'raw_legal_actions': ['call', 'raise']
}, True]
agent.feed(ts)
def test_train(self):
norm_step = 100
memory_init_size = 20
step_num = 1000
sess = tf.InteractiveSession()
tf.Variable(0, name='global_step', trainable=False)
agent = NFSPAgent(sess=sess,
scope='nfsp',
action_num=2,
state_shape=[2],
hidden_layers_sizes=[10, 10],
reservoir_buffer_capacity=50,
batch_size=4,
min_buffer_size_to_learn=memory_init_size,
q_replay_memory_size=50,
q_replay_memory_init_size=memory_init_size,
q_batch_size=4,
q_norm_step=norm_step,
q_mlp_layers=[10, 10])
sess.run(tf.global_variables_initializer())
predicted_action = agent.eval_step({
'obs':
np.random.random_sample((2, )),
'legal_actions': [0, 1]
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
for step in range(step_num):
agent.sample_episode_policy()
predicted_action = agent.step({
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
})
self.assertGreaterEqual(predicted_action, 0)
self.assertLessEqual(predicted_action, 1)
ts = [{
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
},
np.random.randint(2), 0, {
'obs': np.random.random_sample((2, )),
'legal_actions': [0, 1]
}, True]
agent.feed(ts)
if step > norm_step + memory_init_size:
agent.train_rl()
agent.train_sl()
sess.close()
tf.reset_default_graph()
# Count the number of steps
step_counters = [0 for _ in range(env.player_num)]
# Init a Logger to plot the learning curve
logger = Logger(xlabel='timestep',
ylabel='reward',
legend='NFSP on Limit Texas Holdem',
log_path=log_path,
csv_path=csv_path)
for episode in range(episode_num):
# First sample a policy for the episode
for agent in agents:
agent.sample_episode_policy()
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for i in range(env.player_num):
for ts in trajectories[i]:
agents[i].feed(ts)
step_counters[i] += 1
# Train the agent
train_count = step_counters[i] - (memory_init_size + norm_step)
if train_count > 0 and train_count % 64 == 0:
rl_loss = agents[i].train_rl()
sl_loss = agents[i].train_sl()
def train_mahjong():
# Make environment
env = rlcard.make('mahjong', config={'seed': 0})
eval_env = rlcard.make('mahjong', config={'seed': 0})
# Set the iterations numbers and how frequently we evaluate the performance
evaluate_every = 1000
evaluate_num = 1000
episode_num = 10000
# The intial memory size
memory_init_size = 1000
# Train the agent every X steps
train_every = 64
# The paths for saving the logs and learning curves
log_dir = './experiments/mahjong_nfsp_result/'
# Set a global seed
set_global_seed(0)
with tf.Session() as sess:
# Initialize a global step
global_step = tf.Variable(0, name='global_step', trainable=False)
# Set up the agents
agents = []
for i in range(env.player_num):
agent = NFSPAgent(sess,
scope='nfsp' + str(i),
action_num=env.action_num,
state_shape=env.state_shape,
hidden_layers_sizes=[512, 512],
anticipatory_param=0.5,
batch_size=256,
rl_learning_rate=0.00005,
sl_learning_rate=0.00001,
min_buffer_size_to_learn=memory_init_size,
q_replay_memory_size=int(1e5),
q_replay_memory_init_size=memory_init_size,
train_every=train_every,
q_train_every=train_every,
q_batch_size=256,
q_mlp_layers=[512, 512])
agents.append(agent)
random_agent = RandomAgent(action_num=eval_env.action_num)
env.set_agents(agents)
eval_env.set_agents(
[agents[0], random_agent, random_agent, random_agent])
# Initialize global variables
sess.run(tf.global_variables_initializer())
# Init a Logger to plot the learning curvefrom rlcard.agents.random_agent import RandomAgent
logger = Logger(log_dir)
for episode in tqdm(range(episode_num)):
# First sample a policy for the episode
for agent in agents:
agent.sample_episode_policy()
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for i in range(env.player_num):
for ts in trajectories[i]:
agents[i].feed(ts)
# Evaluate the performance. Play with random agents.
if episode % evaluate_every == 0:
logger.log_performance(env.timestep,
tournament(eval_env, evaluate_num)[0])
# Close files in the logger
logger.close_files()
# Plot the learning curve
logger.plot('NFSP')
# Save model
save_dir = 'models/mahjong_nfsp'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
saver = tf.train.Saver()
saver.save(sess, os.path.join(save_dir, 'model'))