decision = random.random()
direction = 0
if decision < epsilon:
direction = random.randint(-1, 1)
else:
shape = (1, ) + last_frames.shape
last_frames_copy = last_frames.reshape(shape)
Q = model.predict(last_frames_copy)
direction = np.argmax(Q) - 1
reward = gameEnv.stepForwardWithReward(direction)
total_reward += reward
over = gameEnv.isOver()
last_frames_new = gameEnv.getLastFrames(num_frames)
print('last_frames shape', last_frames_new.copy().shape)
memoryPiece = [
last_frames.copy(),
last_frames_new.copy(), direction, reward
]
memoryPool.addMemory(memoryPiece)
last_frames = last_frames_new
X_batch, targets = memoryPool.getMemoryBatch(batch_size, model, gamma)
loss += float(model.train_on_batch(X_batch, targets))
if checkpoint_freq and (episode % checkpoint_freq) == 0:
model.save(f'dqn-{episode:08d}.model')
summary = 'Episode {:5d}/{:5d} | Loss {:8.4f} | Total Reward {:4d}'
print(summary.format(episode + 1, num_episode, loss, total_reward))
