def dqn_run_experiments():
for i in range(NUM_EXP):
# Make environment
env = rlcard.make('blackjack', config={'seed': i})
eval_env = rlcard.make('blackjack', config={'seed': i})
# Set the iterations numbers and how frequently we evaluate/save plot
# The initial memory size
memory_init_size = 100
# Train the agent every X steps
train_every = 1
# The paths for saving the logs and learning curves
log_dir = f"{DQN_RES_DIR}/{i}"
# Set up the agents
agent = DQNAgent('dqn',
action_num=env.action_num,
replay_memory_init_size=memory_init_size,
train_every=train_every,
state_shape=env.state_shape,
mlp_layers=[128, 256, 512],
debug=True)
env.set_agents([agent])
eval_env.set_agents([agent])
# Init a Logger to plot the learning curve
logger = Logger(log_dir, debug=True)
for episode in range(DQN_TRAINING_DURATION):
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for ts in trajectories[0]:
agent.feed(ts)
# Evaluate the performance. Play with random agents.
if episode % EVALUATE_EVERY == 0:
logger.log_performance(env.timestep, tournament(eval_env, EVALUATE_NUM_OF_HANDS)[0])
# Close files in the logger
# logger.close_files()
# Plot the learning curve
logger.plot(f"DQN_{i}")
BaseAgent.plot_avg(DQN_RES_DIR, "DQN")
def train(self):
for i in range(0, TRAINING_DURATION // EVALUATE_EVERY + 1):
self.logger.log_performance(
i * EVALUATE_EVERY,
tournament(self._env, EVALUATE_NUM_OF_HANDS)[0])
# Best Alpha so far is 0.015
self.eval_policy, self.q = mc_control(
self._env,
q=self.q,
to_train=EVALUATE_EVERY,
already_trained=EVALUATE_EVERY * i,
alpha=self.alpha,
gamma=1.0,
eps_start=1.0,
eps_decay=0.99999,
eps_min=0.015)
self.policy = self.eval_policy
def play(self, num_plays=NUM_HANDS):
return tournament(self._env, num_plays)
def train(self):
for i in range(0, TRAINING_DURATION // EVALUATE_EVERY + 1):
self.logger.log_performance(
i * EVALUATE_EVERY,
tournament(self._env, EVALUATE_NUM_OF_HANDS)[0])
self.policy = self.eval_policy