def main(args):
# environment
env = GridWorld()
# agent
agent = TDAgent(
env, epsilon=args.epsilon, gamma=args.discout, alpha=0.05, lamda=0.7)
agent.control(method=args.algorithm)
G = Qs[tau % n]
for k in range(tau, min(tau + n - 1, T - 1)):
G += Z * deltas[k % n]
Z = gamma * Z * ((1 - sigma) * pis[(k + 1) % n] + sigma)
p = p * (1 - sigma + sigma * ratios[k % n])
s = states[tau % n]
a = actions[tau % n]
# Update state-action value function.
Q[s, a] += alpha * p * (G - Q[s, a])
action_values = [Q[s, i] for i in range(4)]
policy[s] = np.argmax(action_values)
t += 1
epsilon = decay(epsilon)
if episode % 100 == 0:
print_episode(episode, n_episodes)
print_episode(n_episodes, n_episodes)
return policy
if __name__ == '__main__':
n = 4
alpha = 0.0001
gamma = 1
sigma = 0.5
epsilon = 1
n_episodes = 50000
n_tests = 10
env = GridWorld()
policy = n_step_Q_sigma(env, n, alpha, gamma, sigma, epsilon, n_episodes)
test_policy(env, policy, n_tests)
r_t = env.step(a_t)
# updates
cumulative_reward += r_t * gamma**step
step += 1
locs.append(env.get_agent_loc())
# termination condition
if env.is_terminal():
break
return cumulative_reward, step, locs
'''train
'''
# define env and agent
env = GridWorld(has_bomb=True)
state_dim = env.height * env.width
n_actions = len(ACTIONS)
agent = Agent(dim_input=state_dim, dim_output=n_actions)
# training params
max_steps = 100
gamma = .9
alpha = 0.03
n_epochs = 1000
n_trials = 5
log_return = np.zeros((n_epochs, 2, n_trials))
log_steps = np.zeros((n_epochs, 2, n_trials))
for epoch_id in range(n_epochs):
def main(args):
env = GridWorld()
agent = TDAgent(env, epsilon=args.epsilon, gamma=args.discount, alpha=args.lr)
agent.control(method=args.algorithm)