scores = []
agent = ReinforceAgent(model, actions=ACTIONS, noise=EXPLORE_RATE)
for _ in range(TEST_EPISODES_PER_EPOCH):
env.reset()
agent.reset()
replay = []
done = False
while not done:
action = agent.process(env.state)
_, reward, done, prevState = env.apply(action)
replay.append((prevState, action, reward))
##
states, actions, rewards = zip( *replay )
actions = ACTIONS.toIndex(actions)
trainable.fit(
[
np.array(states),
np.array(actions),
np.array(discountedReturns(rewards, GAMMA))
],
epochs=1, verbose=0
)
##
scores.append(env.score)
Utils.trackScores(scores, metrics)
##################
EXPLORE_RATE = max((0.001, EXPLORE_RATE * EXPLORE_RATE_DECAY))
plotData2file(metrics, 'chart.jpg')
print('[N = %d, M = %d] Start of %d epoch. Explore rate: %.3f' %
(N, M, epoch, exploreRate))
# train model
model.updateTargetModel()
losses = []
for _ in range(TRAIN_EPISODES):
states, actions, rewards, nextStates, nextStateScoreMultiplier = memory.sampleBatch(
batch_size=BATCH_SIZE, maxSamplesFromEpisode=16)
actions = ACTIONS.toIndex(actions)
_, loss = model.fit(states, actions, rewards, nextStates,
nextStateScoreMultiplier * GAMMA)
losses.append(loss)
######
print('Avg. train loss: %.4f' % (np.mean(losses)))
##################
print('Testing...')
scores = Utils.testAgent(
DQNAgent(model, actions=ACTIONS, exploreRate=exploreRate), memory,
TEST_EPISODES)
Utils.trackScores(scores, metrics)
##################
if ((epoch % 10) == 0) and not COLAB_ENV: # debug
Utils.showAgentPlay(DQNAgent(model, actions=ACTIONS, exploreRate=0))
##################
exploreRate = max((0.001, exploreRate * EXPLORE_RATE_DECAY))
plotData2file(metrics, 'chart-%d-%d.jpg' % (
N,
M,
))