def train(agent, args, env, saver):
"""
Train the agent.
"""
logger = Logger(agent, args, saver.save_dir)
# Pre-fill the Replay Buffer
agent.initialize_memory(args.pretrain, env)
#Begin training loop
for episode in range(1, args.num_episodes + 1):
# Begin each episode with a clean environment
env.reset()
# Get initial state
states = env.states
# Gather experience until done or max_steps is reached
for t in range(args.max_steps):
actions = agent.act(states)
next_states, rewards, dones = env.step(actions)
agent.step(states, actions, rewards, next_states)
states = next_states
logger.log(rewards, agent)
if np.any(dones):
break
saver.save_checkpoint(agent, args.save_every)
agent.new_episode()
logger.step(episode, agent)
env.close()
saver.save_final(agent)
logger.graph()
return
def eval(agent, args, env):
"""
Evaluate the performance of an agent using a saved weights file.
"""
logger = Logger(agent, args)
#Begin evaluation loop
for episode in range(1, args.num_episodes + 1):
# Begin each episode with a clean environment
env.reset()
# Get initial state
states = env.states
# Gather experience until done or max_steps is reached
for t in range(args.max_steps):
actions = agent.act(states, eval=True)
next_states, rewards, dones = env.step(actions)
states = next_states
logger.log(rewards, agent)
if np.any(dones):
break
agent.new_episode()
logger.step(episode)
env.close()
return
def notebook_eval_agent(args, env, filename, num_eps=2):
eval_agent = DQN_Agent(env.state_size, env.action_size, args)
eval_saver = Saver(eval_agent.framework, eval_agent, args.save_dir,
filename)
args.eval = True
logger = Logger(eval_agent, args)
for episode in range(3):
env.reset()
state = env.state
for t in range(args.max_steps):
action = eval_agent.act(state)
next_state, reward, done = env.step(action)
state = next_state
logger.log(reward, eval_agent)
if done:
break
eval_agent.new_episode()
logger.step(episode)
def train(agent, args, env, saver):
"""
Train the agent.
"""
logger = Logger(agent, args, saver.save_dir, log_every=50)
# Pre-fill the Replay Buffer
agent.initialize_memory(args.pretrain, env)
#Begin training loop
for episode in range(1, args.num_episodes + 1):
# Begin each episode with a clean environment
done = False
env.reset()
# Get initial state
state = env.state
# Gather experience until done or max_steps is reached
while not done:
action = agent.act(state)
next_state, reward, done = env.step(action)
if done:
next_state = None
agent.step(state, action, reward, next_state)
state = next_state
logger.log(reward, agent)
saver.save_checkpoint(agent, args.save_every)
agent.new_episode()
logger.step(episode, agent.epsilon)
env.close()
saver.save_final(agent)
logger.graph()
return