class AI:
def __init__(self, fname):
lr = 0.0005
self.agent = Agent(gamma=0.99,
epsilon=0.0,
alpha=lr,
input_dims=6,
n_actions=2,
mem_size=60000,
batch_size=64,
epsilon_end=0.0,
fname=fname)
self.observation = []
self.action = 0
self.n_step = 0
self.fname = fname.split("/")[-1]
def episode_start(self, observation):
self.observation = observation
def choose_action(self):
self.action = self.agent.choose_action(self.observation)
return self.action
def step(self, observation_, reward, done):
self.agent.remember(self.observation, self.action, reward,
observation_, int(done))
self.observation = observation_
if self.n_step % 3 == 0:
self.agent.learn()
self.n_step += 1
def episode_end(self):
self.agent.save_model()
def start():
env = gym.make('CartPole-v0')
params = {
'gamma': 0.8,
'epsi_high': 0.9,
'epsi_low': 0.05,
'decay': 500,
'lr': 0.001,
'capacity': 10000,
'batch_size': 64,
'state_space_dim': env.observation_space.shape[0],
'action_space_dim': env.action_space.n
}
agent = Agent(**params)
score = []
mean = []
for episode in range(1000):
s0 = env.reset()
total_reward = 1
for i in range(200):
env.render()
a0 = agent.act(s0)
s1, r1, done, _ = env.step(a0)
if done:
r1 = -1
agent.put(s0, a0, r1, s1)
if done:
break
total_reward += r1
s0 = s1
agent.learn()
score.append(total_reward)
mean.append(sum(score[-100:]) / 100)
print(total_reward)
def main():
#make env and agent
env = gym.make('LunarLander-v2')
agent = Agent(gamma=0.99,
epsilon=1.0,
batch_size=64,
n_actions=4,
eps_end=0.01,
input_dims=[8],
lr=0.0001)
scores, eps_history = [], []
n_games = 500
for i in range(n_games):
score = 0
done = False
observation = env.reset()
while not done:
#ingame
#get action from current view of game (observation)
action = agent.choose_action(observation)
#next frame
observation_, reward, done, info = env.step(action)
score += reward
#store memory
agent.store_transisation(observation, action, reward, observation_,
done)
agent.learn()
#set next stage to current stage
observation = observation_
#append score and eps
scores.append(score)
eps_history.append(agent.epsilon)
#print some nice statements
avg_score = np.mean(scores[-100:])
print(
f'Episode: {i} Score: {score} Average Score: {avg_score} Epsilon: {agent.epsilon}'
)
def OldStuff():
tf.compat.v1.disable_eager_execution()
lr = 0.001
numGames = 10000
session = TriadGameSession()
observation = session.getState()
scores = []
agent = Agent(gamma=0.99,
lr=lr,
epsilon=1.0,
epsilonDec=0.0005,
inputSize=[len(observation)],
numActions=session.getMaxActions(),
memSize=1000000,
batchSize=1024)
for i in range(numGames):
done = False
score = 0
session = TriadGameSession()
observation = session.getState()
while not done:
action = agent.chooseAction(observation)
observationNext, reward, done = session.step(action)
score += reward
agent.store(observation, action, reward, observationNext, done)
observation = observationNext
agent.learn()
scores.append(score)
avgScore = np.mean(scores[-100:])
print('game:', i, 'score %.2f' % score, 'avgScore %.2f' % avgScore,
'epsilon %.2f' % agent.epsilon)
#agent.save()
print('Finished!')
if i % 10 == 0 and i > 0:
avg_score = np.mean(scores[max(0, i - 10):(i + 1)])
print('episode: ', i, 'score: ', score,
' average score %.3f' % avg_score,
'epsilon %.3f' % brain.EPSILON)
else:
print('episode: ', i, 'score: ', score)
eps_history.append(brain.EPSILON)
done = False
observation = env.reset()
score = 0
while not done:
action = brain.chooseAction(observation)
observation_, reward, done, info = env.step(action)
score += reward
brain.storeTransition(observation, action, reward, observation_,
done)
observation = observation_
brain.learn()
scores.append(score)
for i in range(10):
done = False
observation = env.reset()
while not done:
action = brain.chooseAction(observation)
observation_, reward, done, info = env.step(action)
observation = observation_
env.render()
n_actions=4, batch_size=64)
scores = []
eps_history = []
for i in range(1, n_games+1):
done = False
score = 0
obseervation = env.reset()
while not done:
if show:
env.render()
action = agent.choose_action(obseervation)
obseervation_, reward, done, info = env.step(action)
score += reward
agent.remember(obseervation, action, reward, obseervation_, done)
obseervation = obseervation_
agent.learn()
eps_history.append(agent.epsilon)
scores.append(score)
avg_score = np.mean(scores[max(0, i-100):i+1])
print('epsiode', i, 'score ', score, 'avg score', avg_score)
if i % 10 == 0 and i > 0:
agent.save_model()
plt.plot(scores)
plt.plot(eps_history)
plt.legend(['score', 'epsilon'], loc='upper left')
plt.show()
def main():
scores = []
eps_history = []
info_history = []
# Random starting-points:
env = sky.make(random=True,
xi=(301, 650 - 25),
yi=(100, 300 - 25),
width=15,
height=15,
v_initial=14)
# Fixed starting-point:
#env = sky.make(xi=550)
agent = Agent(gamma=gamma,
epsilon=epsilon,
lr=lr,
input_dims=[imput_dimensions],
n_actions=n_actions,
mem_size=mem_size,
batch_size=batch_size,
epsilon_dec=epsilon_dec)
if (load_checkpoint):
agent.load_modes()
for i in range(n_games):
score = 0
done = False
observation = env.reset()
while not done:
'''
one game: ending, when done=True
'''
action = agent.choose_action(observation)
observation_, reward, done, info = env.step(action)
score += reward
agent.store_transition(observation, action, reward, observation_,
int(done))
observation = observation_
agent.learn()
if i % 10 == 0 and i > 0:
avg_score = np.mean(scores[max(0, i - 10):(i + 1)])
print(i, 'episode', info, '|| score:', score,
'| average score: %.3f' % avg_score,
'| epsilon: %.3f' % agent.epsilon, '| training done:',
round(i / n_games, 2))
else:
print(i, 'episode', info, '|| score:', score)
scores.append(score)
eps_history.append(agent.epsilon)
info_history.append(info)
print('training ended with:',
[[el, info_history.count(el)] for el in ('crashed', 'goal')])
if (save_checkpoint):
agent.save_models()
print('[+] model saved')
# -------------------
# Plotting and output
# -------------------
x = [i + 1 for i in range(n_games)]
# First axis: Scores
fig, ax1 = plt.subplots()
color = 'tab:red'
ax1.set_xlabel('Episode')
ax1.set_ylabel('score per Episode', color=color)
ax1.scatter(x, scores, color=color, s=2)
ax1.tick_params(axis='y', labelcolor=color)
# Second axis: epsilon
ax2 = ax1.twinx() # instantiate a second axes that shares the same x-axis
color = 'tab:blue'
ax2.set_ylabel('epsilon',
color=color) # we already handled the x-label with ax1
ax2.plot(x, eps_history, color=color)
ax2.tick_params(axis='y', labelcolor=color)
# Output
fig.tight_layout() # otherwise the right y-label is slightly clipped
plt.savefig(filename)
return env
# Update memory objects with states for each player
if train_networks == True:
memory_1.store_transition(p1_state, p1_action, p1_reward,
p1_state_, int(done))
memory_2.store_transition(p2_state, p2_action, p2_reward,
p2_state_, int(done))
# Train agent DeepQ Networks
# Assign history values so it doesnt break when memcntr is low
history_1, history_2 = None, None
if train_networks == True and memory_1.mem_cntr > batch_size:
history_1, history_2 = None, None
if p1_type == 'Agent':
history_1 = agent_1.learn(
batch_size, memory_1.sample_buffer(batch_size))
if p2_type == 'Agent':
history_2 = agent_2.learn(
batch_size, memory_2.sample_buffer(batch_size))
# Update state for new step
p1_state = p1_state_
p2_state = p2_state_
p1_tot += p1_reward
p2_tot += p2_reward
# Save networks if they are also being trained
# Moved inside game loop, will save every 1000 frames
# Games are getting longer and need to be saved in progress