def choose_action(state):
# print("choose s= ", state)
prob = np.random.rand()
if prob < compute_epsilon():
return explore(state)
else:
return exploit(state)
for i in range(10000):
steps += 1
if env.game_over():
env.reset_game()
s = (int(agent.getGameState()["snake_head_x"]),
int(agent.getGameState()["snake_head_y"]))
print("----------")
print("s = ", s)
action = choose_action(s)
env.act(action)
print(action)
sprim = (0, 0)
if action == 119:
sprim = (s[0], s[1] + 1)
update_w(s, sprim, action)
if action == 97:
sprim = (s[0] + 1, s[1])
update_w(s, sprim, action)
if action == 100:
sprim = (s[0] - 1, s[1])
if __name__ == "__main__":
game = Snake(width=256, height=256)
env = PLE(game,
display_screen=True,
fps=10,
state_preprocessor=process_state)
agent = DQNAgent(env)
agent.load('./save/snake.h5')
env.init()
for e in range(EPISODE):
env.reset_game()
score = 0
state = game.getGameState()
state = sorted(state.items(), key=operator.itemgetter(0))
for i in range(len(state)):
state[i] = state[i][1]
state[2] = len(state[2])
state = np.array([state])
for time_t in range(5000):
action = agent.act(state)
reward = env.act(action)
score += reward
next_state = game.getGameState()
next_state = sorted(next_state.items(), key=operator.itemgetter(0))
for i in range(len(next_state)):
env = PLE(agent, fps=15, force_fps=False, display_screen=True)
env.init()
actions = env.getActionSet()
q_table = {}
alpha = 0.1
gamma = 0.9
while True:
print(q_table)
old_game_state = round_state(agent.getGameState())
if env.game_over():
env.reset_game()
up = q_table.get(tuple(old_game_state.values()) + (119, ), 0)
right = q_table.get(tuple(old_game_state.values()) + (97, ), 0)
left = q_table.get(tuple(old_game_state.values()) + (100, ), 0)
down = q_table.get(tuple(old_game_state.values()) + (115, ), 0)
list = [up, right, left, down]
max_act = max(list)
counter = 0
if max_act == up: