Based on DeepMind publication, I've recreated the environment and I am trying to make the DQN find and converge to an optimal policy. The task of an agent is to learn how to sustainably collect apples (objects), with the regrowth of the apples depending on its spatial configuration (the more apples around, the higher the regrowth). So in short: the agent has to find how to collect as many apples as he can (for collecting an apple he gets a reward of +1), while simultaneously allowing them to regrow, which maximizes his reward (if he depletes the resource too quickly, he looses future reward). The grid-game is visible on the picture below, where the player is a red square, his direction grey, and apple green: enter image description here As given in the publication, I've built a DQN to solve the game. However, regardless of playing with learning rate, loss, exploration rate and its decay, batch size, optimizer, replay buffer, increasing the NN size the DQN does not find an optimal policy pictured below: enter image description here I wonder if there is some mistake in my DQN code (with the similar implementation I've managed to solve OpenAI Gym CartPole task.) Pasting my code below:

class DDQNAgent(RLDebugger):
    def __init__(self, observation_space, action_space):
        # get size of state and action
        self.state_size = observation_space[0]
        self.action_size = action_space
        # hyper parameters
        self.learning_rate = .00025
        self.model = self.build_model()
        self.target_model = self.model
        self.gamma = 0.999
        self.epsilon_max = 1.
        self.epsilon = 1.
        self.t = 0
        self.epsilon_min = 0.1
        self.n_first_exploration_steps = 1500
        self.epsilon_decay_len = 1000000
        self.batch_size = 32
        self.train_start = 64
        # create replay memory using deque
        self.memory = deque(maxlen=1000000)
        self.target_model = self.build_model(trainable=False)

    # approximate Q function using Neural Network
    # state is input and Q Value of each action is output of network
    def build_model(self, trainable=True):
        model = Sequential()
        # This is a simple one hidden layer model, thought it should be enough here,
        # it is much easier to train with different achitectures (stack layers, change activation)
        model.add(Dense(32, input_dim=self.state_size, activation='relu', trainable=trainable))
        model.add(Dense(32, activation='relu', trainable=trainable))
        model.add(Dense(self.action_size, activation='linear', trainable=trainable))
        model.compile(loss='mse', optimizer=RMSprop(lr=self.learning_rate))
        # 1/ You can try different losses. As an logcosh loss is a twice differenciable approximation of Huber loss
        # 2/ From a theoretical perspective Learning rate should decay with time to guarantee convergence
        return model

    # get action from model using greedy policy
    def get_action(self, state):
        if random.random() < self.epsilon:
            return random.randrange(self.action_size)
        q_value = self.model.predict(state)
        return np.argmax(q_value[0])

    # decay epsilon
    def update_epsilon(self):
        self.t += 1
        self.epsilon = self.epsilon_min + max(0., (self.epsilon_max - self.epsilon_min) *
                        (self.epsilon_decay_len - max(0.,
                                 self.t - self.n_first_exploration_steps)) / self.epsilon_decay_len)

    # train the target network on the selected action and transition
    def train_model(self, action, state, next_state, reward, done):

        # save sample <s,a,r,s'> to the replay memory
        self.memory.append((state, action, reward, next_state, done))

        if len(self.memory) >= self.train_start:
            states, actions, rewards, next_states, dones = self.create_minibatch()

            targets = self.model.predict(states)
            target_values = self.target_model.predict(next_states)

            for i in range(self.batch_size):
                # Approx Q Learning
                if dones[i]:
                    targets[i][actions[i]] = rewards[i]
                    targets[i][actions[i]] = rewards[i] + self.gamma * (np.amax(target_values[i]))

            # and do the model fit!
            loss = self.model.fit(states, targets, verbose=0).history['loss'][0]

            for i in range(self.batch_size):
                self.record(actions[i], states[i], targets[i], target_values[i], loss / self.batch_size, rewards[i])

    def create_minibatch(self):
        # pick samples randomly from replay memory (using batch_size)

        batch_size = min(self.batch_size, len(self.memory))
        samples = random.sample(self.memory, batch_size)

        states = np.array([_[0][0] for _ in samples])
        actions = np.array([_[1] for _ in samples])
        rewards = np.array([_[2] for _ in samples])
        next_states = np.array([_[3][0] for _ in samples])
        dones = np.array([_[4] for _ in samples])

        return (states, actions, rewards, next_states, dones)

    def update_target_model(self):

And this is the code which I use to train the model:

from dqn_agent import *
from environment import *

env = GameEnv()
observation_space = env.reset()

agent = DDQNAgent(observation_space.shape, 7)

state_size = observation_space.shape[0]
last_rewards = []
episode = 0
max_episode_len = 1000
while episode < 2100:
    episode += 1
    state = env.reset()
    state = np.reshape(state, [1, state_size])
    #if episode % 100 == 0:
     #   env.render_env()
    total_reward = 0

     step = 0
       gameover = False
    while not gameover:
        step += 1
        #if episode % 100 == 0:
         #   env.render_env()
        action = agent.get_action(state)
        reward, next_state, done = env.step(action)
        next_state = np.reshape(next_state, [1, state_size])
        total_reward += reward
        agent.train_model(action, state, next_state, reward, done)
        state = next_state
        terminal = (step >= max_episode_len)
        if done or terminal:
            gameover = True

    print('episode:', episode, 'cumulative reward: ', total_reward, 'epsilon:', agent.epsilon, 'step', step)

With the model being updated after each episode (episode=1000 steps).

Looking at logs, the agent sometimes tends to achieve very high results more than few times in a row, but always fails to stabilize and the results from episode to episode have an extremely high variance (even after increasing epsilon and running for few thousands of episodes). Looking at my code and the game, do you have any ideas for what might help the algorithm stabilize the results/converge? I've been playing a lot with hyperparameters but nothing gives very significant improvement.

Some parameters on the game & training: Reward: +1 for collecting each apple (green square) Episode: 1000 steps, after 1000 steps or in case the player completely depletes the resource, the game automatically resets. Target model update: after each game termination Hyperparameters can be found in the code above.

Let me know if you have any ideas, happy to share the github repo. Feel free to email me at [email protected]

P.S. I know that this is a similar problem to the one presented below. But I have tried what has been suggested there with no success, hence decided to create another question. DQN cannot learn or converge

EDIT: Added the reward graph (below).

enter image description here

  • $\begingroup$ This looks like a normal reward graph for a DQN training process. The agent learns from its mistakes, and needs to make mistakes in order to do so. Have you tried assessing the agent using a purely greedy policy? Typically you would stop training every so many episodes and assess the agent without a training loop and with epsilon set to zero, If the environment has any randomness, you should assess multiple times to get a mean result. Could you do that and show the graph? If this solves your problem I could write an answer explaining why you need to do this $\endgroup$ Commented Apr 1, 2019 at 18:50
  • $\begingroup$ I have not tried it yet. I did not do that because I did not see that much point in testing it before I saw the algorithm converge. Is the point of stopping training and assessing the agent without a training loop solely to test the performance of an agent at points in training? I am currently running it for 10000 episodes with a larger network to see if it helps it converge. But tomorrow I am gonna try what you said. $\endgroup$
    – macwiatrak
    Commented Apr 1, 2019 at 22:33
  • $\begingroup$ "Is the point of stopping training and assessing the agent without a training loop solely to test the performance of an agent at points in training?". Yes. The point is that the score during training is not a reliable measure because you have deliberately made the agent act non-optimally and randomly. You should absolutely expect large variations in reward for most environments when using trial-and-error exploration. You may also get significant variation of optimal behaviour depending on how precise optimal behaviour needs to be, but ideally you will get convergence at the end. $\endgroup$ Commented Apr 2, 2019 at 6:20
  • $\begingroup$ Thanks! I am doing that currently together with fitting the hyperparameters in a slightly less complicated environment (that takes less time to train and evaluate the performance). So far the results are promising (managed to converge and stabilize it). If you are interested, I am happy to share the hyperparams once I am sure it's working (few days probably). $\endgroup$
    – macwiatrak
    Commented Apr 3, 2019 at 11:02

1 Answer 1


You say that the agent can achieve high results so it appears to be learning something but it fails to do this consistently. Is there are significant drop off in performance at some points?

Is there no improvement in the results if you use Adam instead of RMSProp? I'd be surprised if this was a "difficult" environment that needed a lot of tweaking. I'm curious as to what your reward graph looks like when compared to just "random" behaviour which I think would also survive in this environment.

  • $\begingroup$ The variance of the rewards is very high and what's interesting is that the extraordinary high/low episode rewards tend to go in streaks (i.e. I can often see that very low results happen in a row (around 3-5 times in a row) and same goes for very high results). Not really, surely not a significant difference between Adam and RMSProp. I've added the reward graph to the question description. Let me know what you think of the graph! For now, I've increased epsilon, increased the net and I will be looking at results after 10k episodes. (Will take few days probably). $\endgroup$
    – macwiatrak
    Commented Apr 1, 2019 at 18:32
  • $\begingroup$ Actually, looks like Adam overperforms RMSProp in the long-term, but I am gonna double check it. $\endgroup$
    – macwiatrak
    Commented Apr 2, 2019 at 2:48
  • $\begingroup$ @macwiatrak In my experience, oscillations in training while improving is a symptom of learning rate (and batch size). The other thing I'd want to compare the reward graph to is exploration rate. If there's a significant correlation in their drop offs it can tell you something. $\endgroup$ Commented Apr 2, 2019 at 3:25
  • $\begingroup$ @macwiatrak If this solved your problem can you mark this as the answer? Thanks $\endgroup$ Commented Apr 10, 2019 at 12:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.