def do_cycle(agent, return_observation=False):
env = simulation.Simulation(
[-100, 0, 500, 0, np.random.rand() * np.pi * 2],
np.deg2rad(15),
dt=0.3)
totalreward = 0
action = 1
observation, done = env.step(action)
observations = [env.state]
for _ in range(300):
action = agent.get_action(observation)
nextobservation, done = env.step(action)
reward = calc_reward(observation, nextobservation)
totalreward += reward
if done:
break
agent.update(observation, action, reward, nextobservation)
observation = nextobservation
observations.append(env.state)
if return_observation:
return totalreward, observations
else:
return totalreward
def step(self):
""" Update with a given step size """
self.world.Step(self.timestep,
self.velocityIterations,
self.positionIterations)
for name, agent in self.agents.items():
agent.update()
def on_update(self):
self.player.update()
for a_object in self.objects:
if a_object.in_camera(): a_object.update()
for agent in self.agents:
if agent.in_camera():
agent.update()
def main():
teamA, teamB = initialize()
ball = [WIDTH // 2, HEIGHT // 2]
while True:
team_red = [agent.get_coordinates() for agent in teamA]
team_blue = [agent.get_coordinates() for agent in teamB]
for agent in teamA:
agent.update(team_own=team_red, team_opp=team_blue, ball=ball)
for agent in teamB:
agent.update(team_own=team_blue, team_opp=team_red, ball=ball)
team_red = [agent.get_coordinates() for agent in teamA]
team_blue = [agent.get_coordinates() for agent in teamB]
field.update_positions(team_red, team_blue, ball, canvas)
root.update()
time.sleep(TIMESTEP)
# agent = agent.SARSAAgent(alpha, discount, env, epsilon=0.6)
# agent = agent.QLearningAgent(alpha, discount, env, epsilon=0.6)
# agent = agent.EVSarsaAgent(alpha, discount, env)
# Initialize environment state -----------
env.reset_state()
reward = 0
# Learning -----------
while (True):
env.render(agent)
input ("=== Episode === ") # Uncomment to inspect agent episode-by-episode
while (True):
# input ("== Step == ") # Uncomment to inspect agent step-by-step
# Get current state
state = env.get_state()
# Choose action
action = agent.get_action(state)
# Try out the action
next_state, reward, terminal = env.step(action)
# Update the agent's internal variable
done = agent.update(state, action, reward, next_state, terminal)
env.render(agent)
if done:
env.reset_state()
break
alpha = 0.2
epsilon = 0.5
discount = 0.99
action_space = env.action_space
state_space = env.state_space
#agent = agent.QLearningAgent(alpha, epsilon, discount, action_space, state_space)
agent = agent.EVSarsaAgent(alpha, epsilon, discount, action_space, state_space)
env.render(agent.qvalues)
state = env.get_state()
while (True):
possible_actions = env.get_possible_actions()
action = agent.get_action(state, possible_actions)
next_state, reward, done = env.step(action)
env.render(agent.qvalues)
next_state_possible_actions = env.get_possible_actions()
agent.update(state, action, reward, next_state,
next_state_possible_actions, done)
state = next_state
if done == True:
env.reset_state()
env.render(agent.qvalues)
state = env.get_state()
continue
mainarg.add_argument("--update_tf_board", type=int, default=1000, help="Update the Tensorboard every X steps.")
args = parser.parse_args()
env = gym.make(args.game_name)
if args.mode == "test":
env.monitor.start('./results/' + args.game_name)
print "Generating Test report."
agent = agent.LearningAgent(env, args)
step = 0
with tf.device("/gpu:0"):
for epoch in range(args.epochs):
for train_step in range(args.train_steps):
done = agent.update(step)
step += 1
agent.save_model()
print("Epoch #", epoch, "has finished.")
if args.mode == "test":
env.monitor.close()
gym.upload('./results/' + args.game_name, api_key='sk_ppjQw9T1TYgT1LKJfSG9Q')
blocked_positions, start_pos, default_reward,
scale)
# Agent -------------
alpha = 0.2
epsilon = 0.25
discount = 0.99
action_space = env.action_space
state_space = env.state_space
agent = agent.QLearningAgent(alpha, epsilon, discount, env)
# Learning -----------
env.render(agent)
state = env.get_state()
while (True):
action = agent.get_explore_action(state)
next_state, reward, done = env.step(action)
env.render(agent)
agent.update(state, action, reward, next_state, done)
state = next_state
if done == True:
env.reset_state()
env.render(agent)
state = env.get_state()
continue
