def main(lr, epsilon, gamma, decay_lr, decay_epsilon, modelfile):
seed = 42
np.random.seed(seed)
env = GridWorld()
agent = Agent(env.get_state_dims(),
env.action_size,
lr,
epsilon,
gamma,
decay_lr,
decay_epsilon,
supervisor=True)
exp_name = "gridworld_lr{}_ep{}_gamma{}_decaylr{}_decayep{}_s{}".format(
lr, epsilon, gamma, decay_lr, decay_epsilon, seed)
print_freq = 10000
logger = Logger(print_freq)
for epochs in range(500000):
s, done, trajectory, score, steps = env.reset(), False, [], 0, 0
while not done and steps < 60: #100:
a = agent.Pi(s, env)
sprime, r, done, interrupt = env.step(a)
a2 = a
if not interrupt:
agent.update(s, a, r, sprime, done)
else:
# interrupt service routine will handle it
# do not update Q table's values
action_dict = {
0: Action.Drop1,
1: Action.Drop2,
2: Action.Drop3,
3: Action.Pick1,
4: Action.Pick2,
5: Action.Pick3
}
s = env.ar.get_state() - 1 if env.ar.is_active() else 5
# for easy debugging purposes
a2 = action_dict[s]
trajectory.append([s, a, a2, r, sprime, done])
s = sprime
score += r
steps += 1
if print_traj:
print(trajectory)
logger.update(epochs, score, steps, env)
if epochs % print_freq == print_freq - 1:
logger.log(epochs)
# print(generate_best_trajectory(env, agent))
agent.decay()
f = open(modelfile, "wb")
pickle.dump(agent, f)
f.close()
