state = np.reshape(state, [1, 4])
# decide action
action = agent.act(state)
if action == 1:
dino.jump()
elif action == 2:
dino.lower()
next_state = dino.get_obstacles()
next_state = np.reshape(next_state, [1, 4])
done = dino.is_game_over()
reward = -0.1 if done else 0.1
agent.remember(state, action, reward, next_state, done)
state = next_state
score = dino.get_score()
scores.append(str(score) + '\n')
print("episode: {}/{}, score: {}".format(e, episodes, score))
# train the agent with the experience of the episode
agent.replay(32)
agent.save()
dino.close()
with open('scores.txt', 'w') as file:
file.writelines(scores)
class Gym:
"""'Gym' class: train the AI agent
Attributes:
'args' (ArgumentParser): parser gethering all the Game parameters
'dino' (Dino): Dino controller
'highscore' (tuple of int, (human, AI)): the best score achieved by a human and an AI
'isHuman' (bool): whether a human or an AI is playing the game
't' (int): number of time steps since the beginning of the game
'agent' (AIAgent, default=None): AI agent playing the game
"""
def __init__(self, args):
super(Gym).__init__()
self.args = args
# environment parameters
self.dino = Dino(args)
# game parameters
self.highscore = load_highscore(args.highscore_filename)
self.t = 0
# to play with an AI
self.isHuman = (args.agent == "human")
if not self.isHuman:
self.agent = AIAgent(args, self.dino)
# load saved parameters
if self.args.load_save:
load_agent(self.agent, self.args.save_filename)
# listen to user inputs
self.inputs_list = []
threading.Thread(target=input_thread,
args=(self.inputs_list, )).start()
def step(self):
"""Play one time step in the game.
"""
# take an action
self.agent.choose_action()
# feed the transition information to the agent
self.agent.set_transition()
# update the number of time steps
self.t += 1
def play(self):
"""Play games continuously.
"""
# display command info
display_info(self.dino.get_n_sim(), self.highscore,
self.args.commands_filename)
handle_user_command(self)
# start the first game
self.dino.start()
while True:
# check if the game is not failed
if not self.dino.is_crashed():
if not self.isHuman:
# check if the game is not paused
if not self.dino.is_playing() and self.args.play_bg:
self.dino.game.resume()
# take a step if the AI is playing
if self.dino.is_playing():
self.step()
# otherwise launch a new game
else:
# reset the number of steps
self.t = 0
# current score
score = self.dino.get_score()
# check if the highscore is beaten
human_score = max(score * self.isHuman, self.highscore[0])
ai_score = max(score * (not self.isHuman), self.highscore[1])
# update the highscore
self.highscore = (human_score, ai_score)
update_score(self.highscore, self.args.highscore_filename)
# display command info
display_info(self.dino.get_n_sim(), self.highscore,
self.args.commands_filename)
handle_user_command(self)
if not self.isHuman:
# save the last simulation
self.agent.reset()
else:
self.dino.start()