def main(agent_info, agent_class, steps, filename):
env_class = floating_horsetrack_environment.Environment
rl_glue = RLGlue(env_class, agent_class)
max_steps = steps
step = 0
episode_end = []
cum_reward = 0
agent_info.update({"actions": env_class.actions})
rl_glue.rl_init(agent_info)
while step < max_steps:
rl_glue.rl_start()
is_terminal = False
while not is_terminal and step < max_steps:
reward, state, action, is_terminal = rl_glue.rl_step()
cum_reward += reward
step += 1
if is_terminal:
episode_end.append(step)
rl_glue.rl_cleanup()
save_results(episode_end, len(episode_end), "data/{}".format(filename))
def main(agent_info, agent_class, env_info, env_class, steps, param_info):
# env_class = horsetrack_environment.Environment
rl_glue = RLGlue(env_class, agent_class)
max_steps = steps
max_episodes = 5
step = 0
episodes = 0
episode_end = np.ones(max_episodes) * max_steps
cum_reward = 0
# max_steps = 20000
agent_info.update({"actions": env_class.actions})
rl_glue.rl_init(agent_info, env_info)
while step < max_steps and episodes < max_episodes:
rl_glue.rl_start()
is_terminal = False
while not is_terminal and step < max_steps:
reward, state, action, is_terminal = rl_glue.rl_step()
cum_reward += reward
step += 1
if is_terminal:
episode_end[episodes] = step
episodes += 1
rl_glue.rl_cleanup()
save_results(episode_end, "{}".format(param_info))
def main():
env_class = horsetrack_environment.Environment
agent_class = random_agent.Agent
rl_glue = RLGlue(env_class, agent_class)
num_episodes = 1000
max_steps = 100000
print("\tPrinting one dot for every run: {}".format(num_episodes),
end=' ')
print("total runs to complete.")
total_steps = [0 for _ in range(max_steps)]
for i in range(num_episodes):
rl_glue.rl_init(agent_info={"actions": env_class.actions})
rl_glue.rl_start()
is_terminal = False
while rl_glue.num_steps < max_steps and not is_terminal:
reward, state, action, is_terminal = rl_glue.rl_step()
# optimal_action[num_steps] += 1 if "action is optimal" else 0
total_steps[i] = rl_glue.num_steps
rl_glue.rl_cleanup()
print(".", end='')
sys.stdout.flush()
# prop_optimal = [num_optimal / num_episodes for num_optimal in optimal_action]
save_results(total_steps, len(total_steps), "RL_EXP_OUT.dat")
print("\nDone")
def main(data_output_location="new_data"):
env_class = horsetrack_environment.Environment
agent_class = random_agent.Agent
agent_name = agent_class.__module__[agent_class.__module__.find(".") + 1:]
environment_name = env_class.__module__[env_class.__module__.find(".") +
1:]
rl_glue = RLGlue(env_class, agent_class)
# num_episodes = 2000
# max_steps = 1000
max_total_steps = 100_000
for epsilon in [0.0, 0.1]:
for alpha in [2, 1, 0.5, 0.25, 0.125, 0.0625]:
print("Running Agent: {} on Environment: {}.".format(
agent_name, environment_name))
agent_init_info = {
"actions": [-1, 1],
"world_size": 100,
"epsilon": epsilon,
"alpha": alpha
}
termination_times = []
rl_glue.rl_init(agent_init_info=agent_init_info)
step_counter = 0
while step_counter < max_total_steps:
rl_glue.rl_start()
is_terminal = False
while step_counter < max_total_steps and not is_terminal:
reward, state, action, is_terminal = rl_glue.rl_step()
step_counter += 1
rl_glue.rl_cleanup()
# print(".", end='')
sys.stdout.flush()
if is_terminal:
termination_times.append(step_counter)
epoch_datetime = int(
(datetime.datetime.now() -
datetime.datetime.utcfromtimestamp(0)).total_seconds())
save_results(
termination_times, len(termination_times),
"{}/{}_{}__{}__epsilon{}__alpha{}.dat".format(
data_output_location, epoch_datetime, agent_name,
environment_name, epsilon, alpha))
print("\nDone")
def run_experiment(env_info, agent_info,
num_episodes=5000,
value_error_threshold=1e-8,
plot_freq=10):
env = GridEnvironment
agent = TDAgent
rl_glue = RLGlue(env, agent)
rl_glue.rl_init(agent_info, env_info)
steps = []
for episode in range(1, num_episodes + 1):
rl_glue.rl_episode(0) # no step limit
steps.append(rl_glue.agent.agent_message("get_steps"))
if episode % plot_freq == 0:
values = rl_glue.agent.agent_message("get_values")
print(rl_glue.environment.env_message("get_grid_state"))
rl_glue.rl_cleanup()
values = rl_glue.agent.agent_message("get_values")
return [values, steps]