def load_data_make_train_play():
frame_stack, moves, rewards = load_saved_data()
model = make_a_model(9, 300, 25, 1)
rewards = pong.discount_rewards(rewards, 0.95)
trained_model = train_model(model, frame_stack, moves, rewards, 5, 100)
pong.play_games(10,
21,
single_player=True,
show_moves=True,
game_agent=trained_model,
save_moves=False)
#load_data_make_train_play()
#make_train_and_return_model()
#main()
# frame_stack, moves, rewards=load_saved_data()
# print(len(moves))
# print(frame_stack[0])
# print(frame_stack[100])
# print(frame_stack[200])
# print(frame_stack[300])
# print(frame_stack[400])
# # print(frame_stack[5])
# # print(frame_stack[6])
# # print(frame_stack[7])
# # print(frame_stack[8])
# # print(frame_stack[9])
# # print(frame_stack[10])
# print(moves[0:200])
# # print(len([x for x in rewards if x==1]))
def make_train_and_return_model():
frame_stack, moves, rewards = pong.play_games(5000, 21, show_moves=False, save_moves=True)
model = make_a_model(4, 50, 10, 1)
model = train_model(model, frame_stack, moves, rewards, 10, 200)
model.save("model_saved")
# print(frame_stack[1].reshape(1,9))
# print(model.predict(frame_stack[1].reshape(1,9)))
pong.play_games(10, 20, game_agent=model)
def main_train_evolve():
my_model = make_a_model(3, 64, 32) # my main model
theta = get_model_weights(my_model) # my main weights
std_dev = 0.02 #drop this guy
lr = 0.1 #play with this
generations = 500 # instaed of epoch/episode
n_trials = 50
for generation in range(generations):
print(f"Season {generation}")
noise = []
fitness = []
for n in range(n_trials):
if (n % 20 == 0):
print(n)
noise.append(np.random.randn(len(theta)) * std_dev)
set_model_weights(my_model, noise[-1] + theta) #make negative, double the search space
frames, moves, rewards = pong.play_games(1, 1, game_agent=my_model, save_moves=False, use_turtle=False,
show_moves=False)
fitness.append(sum(rewards))
# print(n, fitness, noise[-1])
ranked_fittness = centered_ranker(np.array(fitness, dtype=float))
print(f"Max fitness {max(fitness)} Average fitness {np.average(fitness)}")
weighted_average_noise = np.dot(ranked_fittness, noise)
theta += weighted_average_noise * lr
set_model_weights(my_model, theta)
my_model.save("model_saved" + str(generation))
print("awe")
def load_model_train_on_play_data_return_results(model_name, games_to_play=100):
model = load_model("model_saved")
frames, moves, rewards = load_saved_data()
model = train_model(model, frames, moves, rewards, 10, 3)
frames, moves, rewards = pong.play_games(10, 10, show_moves=False, save_moves=False,
game_agent=model)
return frames, moves, rewards
def main():
play_random = True
decision_object = None
games_to_play = 2000
points_per_game = 21
single_player = True
frame_stack, moves, rewards = pong.play_games(games_to_play,
points_per_game,
show_moves=False)
def load_data_make_train_play():
frame_stack, moves, rewards = load_saved_data()
model = make_a_model(4, 300, 25, 1)
rewards = pong.discount_rewards(rewards, 0.95)
trained_model = train_model(model, frame_stack, moves, rewards, 5, 100)
pong.play_games(10, 21, single_player=True, show_moves=True, game_agent=trained_model, save_moves=False)
def make_model_play(model_name, games, show_moves, save_moves=False,ball_speed=10):
model = load_model(model_name)
frames, moves, rewards = pong.play_games(games, 10, show_moves=show_moves, save_moves=save_moves, game_agent=model,
ball_speed=100, use_turtle=True)
print(sum(rewards))
return frames, moves, rewards
def generate_random_moves_and_save():
frame_stack, moves, rewards = pong.play_games(10000, 21, single_player=True, show_moves=False, game_agent=None,
save_moves=True, use_turtle=False)
print(len(frame_stack))