def main():
esplison = 0.6
memorydb_instance = MemoryDB('localhost', 'mario-ai', 'replay-memory')
agent_instance = Agent((50, 75, 4), False)
i = 1
while True:
print(f"experiences size: {memorydb_instance.get_experiences_size()}")
sampled_experiences, b_idx, b_ISWeights = memorydb_instance.sample(
sample_size)
print("sampled_experiences: ", len(sampled_experiences))
errors = train_with_experience(agent_instance, sampled_experiences,
b_ISWeights, sample_size, epoch,
discount)
agent_instance.save_model()
memorydb_instance.update_batch(b_idx, errors, sampled_experiences)
if i % training_before_update_target == 0:
agent_instance.sync_target()
i += 1
def main():
esplison = 0.6
io_instance = IO("FCEUX 2.2.3: mario")
memorydb_instance = MemoryDB('localhost', 'mario-ai', 'replay-memory')
agent_instance = Agent((50, 75, 4), False)
i = 1
while True:
print(f"experiences size: {memorydb_instance.get_experiences_size()}")
experience_bacth = []
io_instance.focus_window()
io_instance.reset()
is_termnial = False
previous_screenshot = io_instance.get_screenshot()
previous_device_state = io_instance.get_device_state()
previous_image_state = io_instance.get_stacked_frames(
previous_screenshot, True)
while is_termnial != True:
experience = {}
experience["terminal"] = False
experience["screenshot"] = previous_screenshot
experience["image_state"] = previous_image_state
experience["device_state"] = previous_device_state
dice = random.uniform(0.0, 1.0)
action_index = 0
if dice >= esplison:
#if True:
reward = agent_instance.model_predict([
experience["image_state"].reshape(1, 50, 75, 4),
experience["device_state"].reshape(1, 4)
])
# reward = output[0]
# print("###value: ", output[2])
# print("###advantage: ", output[1])
print("###reward: ", reward)
action_index = np.argmax(reward).item()
print("Model selected action and rewards:", action_index,
io_instance.action_mapping_name[action_index])
else:
action_index = random.randint(0, 3)
print("Random selected action:", action_index,
io_instance.action_mapping_name[action_index])
io_instance.action(action_index)
experience["action_index"] = action_index
experience[
"next_screenshot"] = previous_screenshot = io_instance.get_screenshot(
)
experience[
"next_image_state"] = previous_image_state = io_instance.get_stacked_frames(
previous_screenshot, False)
experience[
"next_device_state"] = previous_device_state = io_instance.get_device_state(
)
experience_bacth.append(experience)
if io_instance.is_termnial(experience["next_screenshot"]):
experience["terminal"] = True
is_termnial = True
calculate_rewards(experience_bacth)
memorydb_instance.add_batch(experience_bacth)
print("Experiences updated")
sampled_experiences, b_idx, b_ISWeights = memorydb_instance.sample(
sample_size)
print("sampled_experiences: ", len(sampled_experiences))
errors = train_with_experience(agent_instance, sampled_experiences,
b_ISWeights, sample_size, epoch,
discount)
agent_instance.save_model()
memorydb_instance.update_batch(b_idx, errors, sampled_experiences)
esplison -= esplison_decay
if i % training_before_update_target == 0:
agent_instance.sync_target()
i += 1
if info["flag_get"]:
reward = 15
new_experience = []
new_experience.append(state)
new_experience.append(action)
new_experience.append(reward)
new_experience.append(done)
new_experience.append(next_state)
memorydb_instance.add(new_experience)
state = next_state
if memorydb_instance.get_experiences_size(
) > experiences_before_training:
sampled_experiences, b_idx, b_ISWeights = memorydb_instance.sample(
sample_size)
errors = agent_instance.train_with_experience(
sampled_experiences, b_ISWeights, epoch, gamma)
agent_instance.save_model()
memorydb_instance.update_batch(b_idx, errors,
sampled_experiences)
score += reward
if done:
# every episode update the target model to be same with model (donkey and carrot), carries over to next episdoe
agent_instance.sync_target()
scores.append(score)
episodes.append(e)
# 'b' is type of marking for plot
pylab.plot(episodes, scores, 'b')
