def test1(self):
print 'Test 1 -- Regular Case'
grid = [['0', '0', '0', '1'], ['0', 'x', '0', '-1'],
['0', '0', '0', '0']]
gw = gridworld.GridWorld(grid, {(0, 3), (1, 3)}, 0.8)
agent = qlearning.QLearningAgent(gw.get_actions,
epsilon=0.2,
alpha=0.5,
gamma=0.9)
# Training
episodes = 5000
for i in range(episodes):
gw.reset((2, 0))
cur_s = gw.get_current_state()
is_done = False
while not is_done:
a = agent.get_action(cur_s)
last_state, action, next_state, reward, is_done = gw.step(a)
agent.learn(last_state, action, next_state, reward, is_done)
cur_s = next_state
# show optimal policy
opt_policy = gw.get_optimal_policy(agent)
gw.display_policy_grid(opt_policy)
gw.display_value_grid(gw.get_values(agent))
gw.display_qvalue_grid(gw.get_qvalues(agent))
def get_agent(name, *args, **kwargs):
if name == 'QLearningAgent':
import qlearning
return qlearning.QLearningAgent(*args, **kwargs)
elif name == 'DQNAgent':
import dqn
return dqn.DQNAgent(*args, **kwargs)
else:
raise Exception('Invalid agent name.')
env = gym.make('CartPole-v0')
if RECORD:
env = wrappers.Monitor(
env,
'/home/vbalogh/git/reinforcement_learning-stormmax/cartpole-experiment-1',
force=True)
def get_actions(state):
return [0, 1]
agent = qlearning.QLearningAgent(get_actions,
epsilon=EPSILON,
alpha=LEARNING_RATE,
gamma=DISCOUNT_FACTOR,
epsilon_decay=EPSILON_DECAY)
history = []
agent, history = train(agent, env, history)
# if RECORD:
# env.monitor.close()
avg_reward = [
numpy.mean(history[i * 100:(i + 1) * 100])
for i in xrange(int(len(history) / 100))
]
f_reward = plt.figure(1)