action = np.random.randint(0, 4)
else:
qvalues = model.predict(currentState)[0]
action = np.argmax(qvalues)
#Updating the Environment
frame, reward, gameOver = env.step(action)
#We need to reshape the frame(2D) to add it to the nextState (4D)
frame = np.reshape(frame, (1, env.nColumns, env.nRows, 1))
nextState = np.append(nextState, frame, axis=3)
nextState = np.delete(nextState, 0, axis=3)
#Remembering new experience and training the AI
DQN.remember([currentState, action, reward, nextState], gameOver)
inputs, targets = DQN.getBatch(model, batchSize)
model.train_on_batch(inputs, targets)
#Updating the score and current state
if env.collected:
nCollected += 1
currentState = nextState
#Updating the epsilon and saving the model
epsilon -= epsilonDecayRate
epsilon = max(epsilon, minLastEpsilon)
if nCollected > maxNCollected and nCollected > 2:
model.save(filePathToSave)
maxNCollected = nCollected
#Taking an action
if np.random.rand() <= epsilon:
action = np.random.randint(0, 3)
else:
qvalues = model.predict(currentState)[0]
action = np.argmax(qvalues)
#Updating the environment
nextState[0], reward, gameOver, _ = env.step(action)
env.render()
totReward += reward
#Remembering new experience, training the AI and updating current state
dqn.remember([currentState, action, reward, nextState], gameOver)
inputs, targets = dqn.getBatch(model, batchSize)
model.train_on_batch(inputs, targets)
currentState = nextState
#Lowering epsilon and displaying the results
epsilon *= epsilonDecayRate
print('Epoch: ' + str(epoch) + ' Epsilon: {:.5f}'.format(epsilon) +
' Total Reward: {:.2f}'.format(totReward))
rewards.append(totReward)
totReward = 0
plt.plot(rewards)
plt.xlabel('Epoch')