Observe and Look Further: Achieving Consistent Performance on Atari

by   Tobias Pohlen, et al.

Despite significant advances in the field of deep Reinforcement Learning (RL), today's algorithms still fail to learn human-level policies consistently over a set of diverse tasks such as Atari 2600 games. We identify three key challenges that any algorithm needs to master in order to perform well on all games: processing diverse reward distributions, reasoning over long time horizons, and exploring efficiently. In this paper, we propose an algorithm that addresses each of these challenges and is able to learn human-level policies on nearly all Atari games. A new transformed Bellman operator allows our algorithm to process rewards of varying densities and scales; an auxiliary temporal consistency loss allows us to train stably using a discount factor of γ = 0.999 (instead of γ = 0.99) extending the effective planning horizon by an order of magnitude; and we ease the exploration problem by using human demonstrations that guide the agent towards rewarding states. When tested on a set of 42 Atari games, our algorithm exceeds the performance of an average human on 40 games using a common set of hyper parameters. Furthermore, it is the first deep RL algorithm to solve the first level of Montezuma's Revenge.


page 1

page 2

page 3

page 4


Agent57: Outperforming the Atari Human Benchmark

Atari games have been a long-standing benchmark in the reinforcement lea...

Jump-Start Reinforcement Learning

Reinforcement learning (RL) provides a theoretical framework for continu...

Multi-task Deep Reinforcement Learning with PopArt

The reinforcement learning community has made great strides in designing...

Challenges to Solving Combinatorially Hard Long-Horizon Deep RL Tasks

Deep reinforcement learning has shown promise in discrete domains requir...

Obstacle Tower: A Generalization Challenge in Vision, Control, and Planning

The rapid pace of research in Deep Reinforcement Learning has been drive...

First return then explore

The promise of reinforcement learning is to solve complex sequential dec...

Improving Policy Optimization with Generalist-Specialist Learning

Generalization in deep reinforcement learning over unseen environment va...

1 Introduction

In recent years, significant advances in the field of deep Reinforcement Learning (RL) have led to artificial agents that are able to reach human-level control on a wide array of tasks such as some Atari 2600 games Bellemare et al. (2015). In many of the Atari games, these agents learn control policies that far exceed the capabilities of an average human player Gruslys et al. (2018); Hessel et al. (2018); Horgan et al. (2018). However, learning human-level policies consistently across the entire set of games remains an open problem.

We argue that an algorithm needs to overcome three key challenges in order to perform well on all Atari games. The first challenge is processing diverse reward distributions. An algorithm must learn stably regardless of reward density and scale. Mnih et al. (2015) showed that clipping rewards to the canonical interval is one way to achieve stability. However, this clipping operation may change the set of optimal policies. For example, the agent no longer differentiates between striking a single pin or all ten pins in Bowling. Hence, optimizing the unaltered reward signal in a stable manner is crucial to achieving consistent performance across games. The second challenge is reasoning over long time horizons, which means the algorithm should be able to choose actions in anticipation of rewards that might be far away. For example, in Montezuma’s Revenge, individual rewards might be separated by several hundred time steps. In the standard -discounted RL setting, this means the algorithm should be able to handle discount factors close to 1. The third and final challenge is efficient exploration of the MDP. An algorithm that explores efficiently is able to discover long trajectories with a high cumulative reward in a reasonable amount of time even if individual rewards are very sparse. While each problem has been partially addressed in the literature, none of the existing deep RL algorithms have been able to address these three challenges at once.

In this paper, we propose a new Deep Q-Network (DQN) Mnih et al. (2015)

style algorithm that specifically addresses these three challenges. In order to learn stably independent of the reward distribution, we use a transformed Bellman operator that reduces the variance of the action-value function. Learning with the transformed operator allows us to process the unaltered environment rewards regardless of scale and density. We prove that the optimal policy does not change in deterministic MDPs and show that under certain assumptions the operator is a contraction in stochastic MDPs (

i.e., the algorithm converges to a fixed point) (see Sec. 3.2). Our algorithm learns stably even at high discount factors due to an auxiliary temporal consistency (TC) loss. This loss prevents the network from prematurely generalizing to unseen states (Sec. 3.3) allowing us to use a discount factor as high as in practice. This extends the effective planning horizon of our algorithm by one order of magnitude when compared to other deep RL approaches on Atari. Finally, we improve the efficiency of DQN’s default exploration scheme by combining the distributed experience replay approach of Horgan et al. (2018) with the Deep Q-learning from Demonstrations (DQfD) algorithm of Hester et al. (2018). The resulting architecture is a distributed actor-learner system that combines offline expert demonstrations with online agent experiences (Sec. 3.4).

We experimentally evaluate our algorithm on a set of 42 games for which we have demonstrations from an expert human player (see Table 5). Using the same hyper parameters on all games, our algorithm exceeds the performance of an average human player on 40 games, the expert player on 34 games, and state-of-the-art agents on at least 28 games. Furthermore, we significantly advance the state-of-the-art on sparse reward games. Our algorithm is the first to complete the first level of Montezuma’s Revenge and it achieves a new top score of 3997 points on Pitfall! without compromising performance on dense reward games and while only using 5 demonstration trajectories.

2 Related work

Reinforcement Learning with Expert Demonstrations (RLED): RLED seeks to use expert demonstrations to guide the exploration process in difficult RL problems. Some early works in this area Atkeson and Schaal (1997); Schaal (1997) used expert demonstrations to find a good initial policy before fine-tuning it with RL. More recent approaches have explicitly combined expert demonstrations with RL data during the learning of the policy or action-value function Chemali and Lazaric (2015); Kim et al. (2013); Piot et al. (2014)

. In these works, expert demonstrations were used to build an imitation loss function (classification-based loss) or max-margin constraints. While these algorithms worked reasonably well in small problems, they relied on handcrafted features to describe states and were not applied to large MDPs. In contrast, approaches using deep neural networks allow RLED to be explored in more challenging RL tasks such as Atari or robotics. In particular, our work builds upon DQfD 

Hester et al. (2018), which used a separate replay buffer for expert demonstrations, and minimized the sum of a temporal difference loss and a supervised classification loss. Another similar approach is Replay Buffer Spiking (RBS) Lipton et al. (2016), wherein the experience replay buffer is initialized with demonstration data, but this data is not kept for the full duration of the training. In robotics tasks, similar techniques have been combined with other improvements to successfully solve difficult exploration problems Nair et al. (2017); Večerík et al. (2017).

Deep Q-Networks (DQN): DQN Mnih et al. (2015) used deep neural networks as function approximators to apply RL to Atari games. Since that work, many extensions that significantly improve the algorithm’s performance have been developed. For example, DQN uses a replay buffer to store off-policy experiences and the algorithm learns by sampling batches uniformly from the replay buffer; instead of using uniform samples, Schaul et al. (2015) proposed prioritized sampling where transitions are weighted by their absolute temporal difference error. This concept was further improved by Ape-X DQN Horgan et al. (2018) which decoupled the data collection and the learning processes by having many actors feed data to a central prioritized replay buffer that an independent learner can sample from.

Durugkar and Stone (2017) observed that due to over-generalization in DQN, updates to the value of the current state also have an adverse effect on the values of the next state. This can lead to unstable learning when the discount factor is high. To counteract this effect, they constrained the TD update to be orthogonal to the direction of maximum change of the next state. However, their approach only worked on toy domains such as Cart-Pole. Finally, van Hasselt et al. (2016a) successfully extended DQN to process unclipped rewards with an algorithm called PopArt, which adaptively rescales the targets for the value network to have zero mean and unit variance.

3 Algorithm

In this section, we describe our algorithm, which consists of three components: (1) The transformed Bellman operator; (2) The temporal consistency (TC) loss; (3) Combining Ape-X DQN and DQfD.

3.1 DQN Background

Let be a finite, discrete-time MDP where is the state space, the action space, the reward function which represents the one-step reward distribution of doing action in state , the discount factor and a stochastic kernel modelling the one-step Markovian dynamics (

is the probability of transitioning to state

by choosing action in state ). The quality of a policy is determined by the action-value function

where is the expectation over the distribution of the admissible trajectories obtained by executing the policy starting from state and taking action . The goal is to find a policy that maximizes the state-value for all states , i.e., find such that V for all . While there may be several optimal policies, they all share a common optimal action-value function  Puterman (1994). Furthermore, acting greedily with respect to the optimal action-value function yields an optimal policy. In addition, is the unique fixed point of the Bellman optimality operator defined as

for any . Because is a -contraction, we can learn using a fixed point iteration. Starting with an arbitrary function and then iterating for generates a sequence of functions that converges to .

DQN Mnih et al. (2015) is an online-RL algorithm using a deep neural network with parameters as a function approximator of the optimal action-value function . The algorithm starts with a random initialization of the network weights and then iterates


where the expectation is taken with respect to a random sample of states and actions and is the Huber loss Huber (1964) defined as

In practice, the minimization problem in (1

) is only approximately solved by performing a finite and fixed number of stochastic gradient descent (SGD) steps

111Mnih et al. (2015) refer to the number of SGD iterations as target update period. and all expectations are approximated by sample averages.

3.2 Transformed Bellman Operator

Mnih et al. (2015) have empirically observed that the errors induced by the limited network capacity, the approximate finite-time solution to (1), and the stochasticity of the optimization problem can cause the algorithm to diverge if the variance of the optimization target is too high. In order to reduce the variance, they clip the reward distribution to the interval . While this achieves the desired goal of stabilizing the algorithm, it significantly changes the set of optimal policies. For example, consider a simplified version of Bowling where an episode only consists of a single throw. If the original reward is the number of hit pins and the rewards were clipped, any policy that hits at least a single pin would be optimal under the clipped reward function. Instead of reducing the magnitude of the rewards, we propose to focus on the action-value function instead. We use a function that reduces the scale of the action-value function. Our new operator is defined as

Proposition 3.1.

Let be the fixed point of and , then

  1. If for , then .

  2. If is strictly monotonically increasing and the MDP is deterministic (i.e.,  and are point measures for all ), then .


(i) is equivalent to linearly scaling the reward by a constant , which implies the proposition. For (ii) let be the fixed point of and note that where the last equality only holds if the MDP is deterministic. ∎

Proposition 3.1 shows that in the basic cases when either is linear or the MDP is deterministic, has the unique fixed point . Hence, if is an invertible contraction and we use instead of in the DQN algorithm, the variance of our optimization target decreases while still learning an optimal policy. In our algorithm, we use with where the additive regularization term ensures that is Lipschitz continuous (see Proposition A.1). We chose this function because it has the desired effect of reducing the scale of the targets while being Liptschitz continuous and admitting a closed form inverse.

In practice, DQN minimizes the problem in (1) by sampling transitions of the form from a replay buffer where , and . Let be transitions from the buffer with normalized priorities , then for the loss function in (1) using the operator is approximated as

where for DQN and for Double DQN van Hasselt et al. (2016b).

3.3 Temporal consistency (TC) loss

The stability of DQN, which minimizes the TD-loss , is primarily determined by the target . While the transformed Bellman operator provides an atemporal reduction of the target’s scale and variance, instability can still occur as the discount factor approaches 1. Increasing the discount factor decreases the temporal difference in value between non-rewarding states. In particular, unwanted generalization of the neural network to the next state (due to the similarity of temporally adjacent target values) can result in catastrophic TD backups. We resolve the problem by adding an auxiliary temporal consistency (TC) loss of the form


is the current iteration. The TC-loss penalizes weight updates that change the next action-value estimate

. This makes sure that the updated estimates adhere to the operator and thus are consistent over time.

3.4 Ape-X DQfD


Minimize TD loss and imitation loss

Actor Replay Buffer

Agent experience

Expert Replay Buffer

Expert experience






Demonstration data

Agent transitions and initial priorities

Network weights

Expert transitions

Prioritized samples

Prioritized samples

Updated priorities

Updated priorities

Training batch

Replay Buffer

Agent experience


Minimize TD loss




Agent transitions and initial priorities

Network weights

Updated priorities

Prioritized samples
(a) Ape-X DQN (b) Ape-X DQfD (ours)
Figure 1: The figure compares our architecture (b) to the one proposed by Horgan et al. (2018) (a).

In this section, we describe how we combine the transformed Bellman operator and the TC loss with the DQfD algorithm Hester et al. (2018) and distributed prioritized experience replay Horgan et al. (2018). The resulting algorithm, which we call Ape-X DQfD following Horgan et al. (2018), is a distributed DQN algorithm with expert demonstrations that is robust to the reward distribution and can learn at discount factors an order of magnitude higher than what was possible before (i.e.,  instead of ). Our algorithm consists of three components: (1) replay buffers; (2) actor processes; and (3) a learner process. Fig. 1 shows how our architecture compares to the one used by Horgan et al. (2018).

Replay buffers. Following Hester et al. (2018), we maintain two replay buffers: an actor replay buffer and an expert replay buffer. Both buffers store 1-step and 10-step transitions and are prioritized Schaul et al. (2015). The transitions in the actor replay buffer come from actor processes that interact with the MDP. In order to limit the memory consumption of the actor replay buffer, we regularly remove transitions in a FIFO-manner. The expert replay buffer is filled once offline before training commences.

Actor processes. Horgan et al. (2018) showed that we can significantly improve the performance of DQN with prioritized replay buffers by having many actor processes. We follow their approach and use actor processes. Each actor follows an -greedy policy based on the current estimate of the action-value function. The noise levels are chosen as where . Notably, this exploration is closer to the one used by Hester et al. (2018) and is much lower (i.e., less random exploration) than the schedule used by Horgan et al. (2018).

Learner process. The learner process samples experiences from the two replay buffers and minimizes a loss in order to approximate the optimal action-value function. Following Hester et al. (2018), we combine the TD-loss with a supervised imitation loss. Let be transitions of the form with normalized priorities where is 1 if the transition is part of the best (i.e., highest episode return) expert episode and 0 otherwise. The imitation loss is a max-margin loss of the form


where is the margin and is 1 if and 0 otherwise. Combining the imitation loss with the TD loss and the TC loss yields the total loss formulation

Algo. 1, provided in the appendix, shows the entire learner procedure. Note that while we only apply the imitation loss on the best expert trajectory, we still use all expert trajectories for the other two losses.

Our learning algorithm differs from the one used by Hester et al. (2018) in three important ways. First, we do not have a pre-training phase where we minimize only using expert transitions. We learn with a mix of actor and expert transitions from the beginning. Second, we maintain a fixed ratio of actor and expert transitions. For each SGD step, our training batch consists of 75% agent transitions and 25% expert transitions. The ratio is constant throughout the entire learning process. Finally, we only apply the imitation loss to the best expert episode instead of all episodes.






Ape-X DQfD

Ape-X DQfD




Avg. Human

Best Expert


Rainbow DQN  31 / 42 9 / 42  10 / 42 7 / 42  41 / 42 32 / 42 24 / 42
DQfD  11 / 42 7 / 42  11 / 42 2 / 42  40 / 42 25 / 42 13 / 42
Ape-X DQN  34 / 42 35 / 42   28 / 42 15 / 42  40 / 42 31 / 42 31 / 42
Ape-X DQfD  32 / 42 39 / 42 15 / 42  9 / 42  40 / 42 39 / 42 32 / 42
Ape-X DQfD (deeper)  36 / 42 40 / 42 28 / 42  33 / 42   42 / 42 40 / 42 34 / 42
Table 1: The table shows on which fraction of the tested games one approach performs at least as well as the other. The scores used for the comparison are using the no-op starts regime. As described in Sec. 4.1, we compare the agents’ scores to the scores obtained by an average human player and an expert player. Ape-X DQfD (deeper) out-performs the average human on 40 of 42 games.
  No-op starts   Human starts
  Mean Median   Mean Median
Algorithm  42 Games 57 Games 42 Games 57 Games  42 Games 57 Games 42 Games 57 Games
Rainbow DQN  1022% 874% 231% 231%  897% 776% 159% 153%
DQfD  364% 113%  
Ape-X DQN  1770% 1695% 421% 434%  1651% 1591% 354% 358%
Ape-X DQfD  1536% 339%   1461% 302%
Ape-X DQfD (deeper)  2346% 702%   2028% 547%
Table 2: The table shows the human-normalized performance of our algorithm and the baselines. For each game, we normalize the score as and then aggregate over all games (mean or median). Because we only have demonstrations on 42 out of the 57 games, we report the performances on 42 games and also 57 games for baselines not using demonstrations.

4 Experimental evaluation

We evaluate our approach on the same subset of 42 games from the Arcade Learning Environment (ALE) Bellemare et al. (2015) used by Hester et al. (2018). We report the performance using the no-op starts and the human starts test regimes Mnih et al. (2015). The full evaluation procedure is detailed in Sec. C.

4.1 Benchmark results

We compare our approach to Ape-X DQN Horgan et al. (2018), on which our actor-learner architecture is based, DQfD Hester et al. (2018), which introduced the expert replay buffer and the imitation loss, and Rainbow DQN Hessel et al. (2018), which combines all major DQN extensions from the literature into a single algorithm. Note that the scores reported in Horgan et al. (2018) were obtained by running 360 actors. Due to resource constraints, we limit the number of actors to 128 for all Ape-X DQfD experiments. Besides comparing our performance to other RL agents, we are also interested in comparing our scores to a human player. Because our demonstrations were gathered from an expert player, the expert scores are mostly better than the level of human performance reported in the literature Mnih et al. (2015); Wang et al. (2016). Hence, we treat the historical human scores as the performance of an average human and the scores of our expert as expert performance.

We first analyse the performance of the standard dueling DQN architecture Wang et al. (2016) that is also used by the baselines. We report the scores as Ape-X DQfD in Tables 2 and 2. We designed the algorithm to achieve higher consistency over a broad range of games and the scores shown in Table 2 reflect that goal. Whereas previous approaches outperformed an average human on at most 32 out of 42 games, Ape-X DQfD with the standard dueling architecture achieves a new state-of-the-art result of 39 out of 42 games. This means we significantly improve the performance on the tails of the distribution of scores over the games. When looking at this performance in the context of the median human-normalized scores reported in Table 2, we see that we significantly increase the set of games where we learn good policies at the expense of achieving lower peak scores on some games.

One of the significant changes in our experimental setup is moving from a discount factor of to . Jiang et al. (2015) argue that this increases the complexity of the learning problem and, thus, requires a bigger hypothesis space. Hence, in addition to the standard architecture, we also evaluated a slightly wider (i.e., double the number of convolutional kernels) and deeper (one extra fully connected layer) network architecture (see Fig. 8). With the deeper architecture, our algorithm outperforms an average human on 40 out of 42 games. Furthermore, it is the first deep RL algorithm to learn non-trivial policies on all games including sparse reward games such as Montezuma’s Revenge, Private Eye, and Pitfall!. For example, we achieve 3997 points in Pitfall!, which is below the 6464 points of an average human but far above any baseline. Finally, with a median human-normalized score of 702% and exceeding every baseline on at least of the games, we demonstrate strong peak performance and consistency over the entire benchmark.

4.2 Imitation vs. inspiration

Figure 2: The figure shows the cumulative undiscounted episode return over time and compares the best expert episode to the best Ape-X DQfD episode on three games. On Hero, the algorithm exceeds the expert’s performance, on Montezuma’s Revenge, it matches the expert’s score but reaches it faster, and on Ms. Pacman, the expert is still superior.
Figure 3: Results of our ablation study using the standard network architecture. The experiment without expert data (

) was performed with the higher exploration schedule used in Horgan et al. (2018).
Figure 2: The figure shows the cumulative undiscounted episode return over time and compares the best expert episode to the best Ape-X DQfD episode on three games. On Hero, the algorithm exceeds the expert’s performance, on Montezuma’s Revenge, it matches the expert’s score but reaches it faster, and on Ms. Pacman, the expert is still superior.

Although we use demonstration data, the goal of RLED algorithms is still to learn an optimal policy that maximizes the expected -discounted return. While Table 2 shows that we exceed the best expert episode on 34 games using the deeper architecture, it is hard to grasp the qualitative differences between the expert policies and our algorithm’s policies. In order to qualitatively compare the agent and the expert, we provide videos on YouTube (see Sec. F) and we plot the cumulative episode return of the best expert and agent episodes in Fig. 3. We see that our algorithm (

) finds more time-efficient policies than the expert (

) in all cases. This is a strong indicator that our algorithm does not do pure imitation but improves upon the demonstrated policies.

4.3 Ablation study

We evaluate the performance contributions of the three key ingredients of Ape-X DQfD (transformed Bellman operator, the TC-loss, and demonstration data) by performing an ablation study on a subset of 6 games. We chose sparse-reward games (Montezuma’s Revenge, Private Eye), dense-reward games (Ms. Pacman, Seaquest), and games where DQfD performs well (Hero, Kangaroo) (see Fig. 3).

Transformed Bellman operator (

When using the standard Bellman operator instead of the transformed one, Ape-X DQfD is stable but the performance is significantly worse.

TC loss (

In our setup, the TC loss is paramount to learning stably. We see that without the TC loss the algorithm learns faster at the beginning of the training process. However, at some point during training, the performance collapses and often the process dies with floating point exceptions.

Expert demonstrations (


Unsurprisingly, removing demonstrations entirely (

) severely degrades the algorithm’s performance on sparse reward games. However, in games that an -greedy policy can efficiently explore, such as Seaquest, the performance is on par or worse. Hence, the bias induced by the expert data is beneficial in some games and harmful in others. Just removing the imitation loss (

) does not have a significant effect on the algorithm’s performance. This stands in contrast to the original DQfD algorithm and is most likely because we only apply the loss on a single expert trajectory.

4.4 Comparison to related work

Figure 4: The figures show how our algorithm compares when we substitute the transformed Bellman operator to PopArt and when we substitute the TC loss to constrained TD updates. Note that the scales differ from the ones in Fig. 3 because the experiments only ran for 40 hours.

The problems of handling diverse reward distributions and network over-generalization in deep RL have been partially addressed in the literature (see Sec. 2). Specifically, van Hasselt et al. (2016a) proposed PopArt and Durugkar and Stone (2017) used constrained TD updates. We evaluate the performance of our algorithm when using alternative solutions and report the results in Fig. 4.

PopArt (

We use the standard Bellman operator in combination with PopArt, which adaptively normalizes the targets in (1

) to have zero mean and unit variance. While the modified algorithm manages to learn in some games, the overall performance is significantly worse than Ape-X DQfD. One possible limiting factor that makes PopArt a bad choice for our framework is that training batches contain highly rewarding states from the very beginning of training. SGD updates performed before the moving statistics have adequately adapted the moments of the target distribution might result in catastrophic changes to the network’s weights.

Constrained TD updates (

We replaced the TC-loss with the constrained TD update approach Durugkar and Stone (2017) that removes the target network and constrains the gradient to prevent an SGD update from changing the predictions at the next state. We did not find the approach to work in our case.

5 Conclusion

In this paper, we presented a deep Reinforcement Learning (RL) algorithm that achieves human-level performance on a wide variety of MDPs on the Atari 2600 benchmark. It does so by addressing three challenges: handling diverse reward distributions, acting over longer time horizons, and efficiently exploring on sparse reward tasks. We introduce novel approaches for each of these challenges: a transformed Bellman operator, a temporal consistency loss, and a distributed RLED framework for learning from human demonstrations and task reward. Our algorithm exceeds the performance of an average human on 40 out of 42 Atari 2600 games and it is the first deep RL algorithm to complete the first level of Montezuma’s Revenge.


  • Atkeson and Schaal [1997] Christopher Atkeson and Stefan Schaal. Robot learning from demonstration. In Proc. of ICML, 1997.
  • Bellemare et al. [2015] Marc Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The Arcade Learning Environment: An evaluation platform for general agents. In Proc. of IJCAI, 2015.
  • Chemali and Lazaric [2015] Jessica Chemali and Alessandro Lazaric. Direct policy iteration with demonstrations. In Proc. of IJCAI, 2015.
  • Durugkar and Stone [2017] Ishan Durugkar and Peter Stone. TD learning with constrained gradients. In Deep Reinforcement Learning Symposium, NIPS, 2017.
  • Gruslys et al. [2018] Audrunas Gruslys, Will Dabney, Mohammad Gheshlaghi Azar, Bilal Piot, Marc Bellemare, and Remi Munos. The reactor: A fast and sample-efficient actor-critic agent for reinforcement learning. In Proc. of ICLR, 2018.
  • Hessel et al. [2018] Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will Dabney, Dan Horgan, Bilal Piot, Mohammad G. Azar, and David Silver. Rainbow: Combining improvements in deep reinforcement learning. In Proc. of AAAI, 2018.
  • Hester et al. [2018] Todd Hester, Matej Vecerik, Olivier Pietquin, Marc Lanctot, Tom Schaul, Bilal Piot, Andrew Sendonaris, Gabriel Dulac-Arnold, Ian Osband, and John Agapiou. Deep Q-learning from demonstrations. Proc. of AAAI, 2018.
  • Horgan et al. [2018] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. In Proc. of ICLR, 2018.
  • Huber [1964] Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73–101, 03 1964.
  • Jiang et al. [2015] Nan Jiang, Alex Kulesza, Satinder Singh, and Richard Lewis. The dependence of effective planning horizon on model accuracy. In Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems, pages 1181–1189. International Foundation for Autonomous Agents and Multiagent Systems, 2015.
  • Kim et al. [2013] Beomjoon Kim, Amir-massoud Farahmand, Joelle Pineau, and Doina Precup. Learning from limited demonstrations. In Proc. of NIPS, 2013.
  • Lipton et al. [2016] Zachary Lipton, Xiujun Li, Jianfeng Gao, Lihong Li, Faisal Ahmed, and Li Deng. Bbq-networks: Efficient exploration in deep reinforcement learning for task-oriented dialogue systems. In Proc. of AAAI, 2016.
  • Mnih et al. [2015] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
  • Nair et al. [2017] Ashvin Nair, Bob McGrew, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. Overcoming exploration in reinforcement learning with demonstrations. arXiv preprint arXiv:1709.10089, 2017.
  • Piot et al. [2014] Bilal Piot, Matthieu Geist, and Olivier Pietquin. Boosted bellman residual minimization handling expert demonstrations. In Proc. of ECML/PKDD, 2014.
  • Puterman [1994] Marc L. Puterman. Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons, 1994.
  • Schaal [1997] Stefan Schaal. Learning from demonstration. In Proc. of NIPS, 1997.
  • Schaul et al. [2015] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. In Proc. of ICLR, 2015.
  • van Hasselt et al. [2016a] Hado van Hasselt, Arthur Guez, Matteo Hessel, Volodymyr Mnih, and David Silver. Learning values across many orders of magnitude. In Proc. of NIPS, 2016a.
  • van Hasselt et al. [2016b] Hado van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double Q-learning. In Proc. of AAAI, 2016b.
  • Večerík et al. [2017] Matej Večerík, Todd Hester, Jonathan Scholz, Fumin Wang, Olivier Pietquin, Bilal Piot, Nicolas Heess, Thomas Rothörl, Thomas Lampe, and Martin Riedmiller. Leveraging demonstrations for deep reinforcement learning on robotics problems with sparse rewards. arXiv preprint arXiv:1707.08817, 2017.
  • Wang et al. [2016] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado van Hasselt, Marc Lanctot, and Nando de Freitas. Dueling network architectures for deep reinforcement learning. In Proc. of ICML, 2016.

Appendix A Transformed Bellman Operator in Stochastic MDPs

The following proposition shows that transformed Bellman operator is still a contraction for small if we assume a stochastic MDP and a more generic choice of . However, the fixed point might not be .

Proposition A.1.

Let be strictly monotonically increasing, Lipschitz continuous with Lipschitz constant , and have a Lipschitz continuous inverse with Lipschitz constant . For , is a contraction.


Let be arbitrary. It holds

where we used Jensen’s inequality in (1) and the Lipschitz properties of and in (1) and (2). ∎

For our algorithm, we use with . While Proposition A.2 shows that the transformed operator is a contraction, the discount factor we use in practice is higher than . We leave a deeper investigation of the contraction properties of in stochastic MDPs for future work.

Proposition A.2.

Let and . It holds

  1. is strictly monotonically increasing.

  2. is Lipschitz continuous with Lipschitz constant .

  3. is invertible with .

  4. is strictly monotonically increasing.

  5. is Lipschitz continuous with Lipschitz constant .

We use the following Lemmas in order to prove Proposition A.2.

Lemma A.1.

is differentiable everywhere with derivative for all .

Proof of Lemma a.1.

For , is differentiable as a composition of differentiable functions with . Analogously, is differentiable for with . For , we find

and similarly . Hence, for all . ∎

Lemma A.2.

is differentiable everywhere with derivative for all .

Proof of Lemma a.2.

For , is differentiable as a composition of differentiable functions. For , it holds

and similarly , which concludes the proof. ∎

Proof of Proposition a.2.

We prove all statements individually.

  1. for all , which implies the proposition.

  2. Let with , using the mean value theorem, we find

  3. (i) Implies that is invertible and simple substitution shows .

  4. for all , which implies the proposition.

  5. Let with , using the mean value theorem, we find

Appendix B Learner algorithm

Random sample
for  do
     for  do is the target network update period
          SamplePrioritized(N) Sample 75% agent and 25% expert transitions
           AdamStep() Update the parameters using Adam
           Compute the updated priorities based on the TD error
          UpdatePriorities() Send the updated priorities to the replay buffers
     end for
end for
Algorithm 1 The algorithm used by the learner to estimate the action-value function.

Appendix C Experimental setup

We evaluate our algorithm on Arcade Learning Environment (ALE) by Bellemare et al. [2015]. While we follow many of the practices commonly applied when training on the ALE, our experimental setup differs in a few key aspects from the defaults Mnih et al. [2015], Hessel et al. [2018].

End episode on life loss.

Most authors who train agents on the ALE choose to end a training episode when the agent loses a life. This naturally makes the agent risk averse as an action that leads to the termination of an episode has a value of 0. However, because our expert player was allowed to continue playing an episode after losing a life, we follow Hester et al. [2018] and only terminate a training episode either when the game is over or when the agent has performed 50,000 steps which is the episode length used by Horgan et al. [2018].

Reward Preprocessing.

As explained in Sec. 3.2, we do not clip the rewards to the interval . Instead, we use the raw and unprocessed rewards provided by each game.

Discount factor.

The majority of approaches use a discount factor of . Empirically, this used to be the highest discount factor that allows stable learning on all games. However, the TC loss allows us to use a much higher discount factor of giving the algorithm an effective planning horizon of 1000 instead of 100 steps.

Expert data.

Instead of relying purely on an -greedy exploration strategy, our algorithm uses expert demonstrations. By using these demonstrations in the TD loss , the algorithm gets the experience rewarding transitions without having to discover them itself.

Appendix D Full experimental results

Figure 5: Training curves on all 42 games. We report the performance using the standard network architecture Wang et al. [2016] and the slightly deeper version (see Fig. 8).
Game  Rainbow DQfD Ape-X DQN  Ape-X DQfD Ape-X DQfD (deeper)  Random Avg. Human Expert
Alien  9491.7 4737.5 40804.9  11313.6 50113.6   128.3 7128.0 29160.0
Amidar  5131.2 2325.0 8659.2  8463.8 12291.7   11.8 1720.0 2341.0
Assault  14198.5 1755.7 24559.4  22855.0 35046.9   166.9 742.0 2274.0
Asterix  428200.3 5493.6 313305.0  399888.0 418433.5  164.5 8503.0 18100.0
Asteroids  2712.8 3796.4 155495.1   116846.4 112573.6  871.3 47389.0 18100.0
Atlantis  826659.5 920213.9 944497.5  911025.0 1057521.5   13463.0 29028.0 22400.0
Bank Heist  1358.0 1280.2 1716.4  2061.9 2578.9  21.7 753.0 7465.0
Battle Zone  62010.0 41708.2 98895.0  60540.0 128925.0   3560.0 37188.0 60000.0
Beam Rider  16850.2 5173.3 63305.2  47129.4 87257.4   254.6 16926.0 19844.0
Bowling  30.0 97.0 17.6  216.3 210.9  35.2 161.0 149.0
Boxing  99.6 99.1 100.0   100.0 98.5  -0.1 12.0 15.0
Breakout  417.5 308.1 800.9   419.7 641.9  1.6 30.0 79.0
Chopper Command  16654.0 6993.1 721851.0  96653.0 840023.5   644.0 7388.0 11300.0
Crazy Climber  168788.5 151909.5 320426.0   176598.5 247651.0  9337.0 35829.0 61600.0
Defender  55105.0 27951.5 411943.5   51442.0 218006.3  1965.5 18689.0 18700.0
Demon Attack  111185.2 3848.8 133086.4  100200.9 141444.6   208.3 1971.0 6190.0
Double Dunk  -0.3 -20.4 23.5   23.0 23.2  -16.0 -16.0 -14.0
Enduro  2125.9 1929.8 2177.4   1663.1 1910.1  81.8 860.0 803.0
Fishing Derby  31.3 38.4 44.4  66.1 68.0   -77.1 -39.0 20.0
Freeway  34.0 31.4 33.7  32.0 31.7  0.1 30.0 32.0
Gopher  70354.6 7810.3 120500.9   114702.6 114168.9  250.0 2412.0 22520.0
Gravitar  1419.3 1685.1 1598.5  4214.3 3920.5  245.5 3351.0 13400.0
Hero  55887.4 105929.4 31655.9  112042.4 114248.2   1580.3 30826.0 99320.0
Ice Hockey  1.1 -9.6 33.0   3.4 32.9  -9.7 1.0 1.0
James Bond  19809.0 2095.0 21322.5   12889.0 16956.3  33.5 303.0 650.0
Kangaroo  14637.5 14681.5 1416.0  47676.5 48599.0   100.0 3035.0 36300.0
Krull  8741.5 9825.3 11741.4  104160.3 140670.6   1151.9 2666.0 13730.0
Kung Fu Master  52181.0 29132.0 97829.5  67957.5 137804.5   304.0 22736.0 25920.0
Montezuma’s Revenge  384.0 4638.4 2500.0  29384.0 27926.5  25.0 4753.0 34900.0
Ms. Pacman  5380.4 4695.7 11255.2  12857.1 20872.7  197.8 6952.0 55021.0
Name This Game  13136.0 5188.3 25783.3  24465.8 31569.4   1747.8 8049.0 19380.0
Pitfall!  0.0 57.3 -0.6  3996.7 3997.5  -348.8 6464.0 47821.0
Pong  20.9 10.7 20.9  21.0 20.9  -18.0 15.0 0.0
Private Eye  4234.0 42457.2 49.8  100747.4 100724.9  662.8 69571.0 72800.0
Q*bert  33817.5 21792.7 302391.3   71224.4 91603.5  183.0 13455.0 99450.0
Riverraid  22920.8 18735.4 63864.4   24147.7 47609.9  588.3 17118.0 39710.0
Road Runner  62041.0 50199.6 222234.5  507213.0 578806.5   200.0 7845.0 20200.0
Seaquest  15898.9 12361.6 392952.3   13603.8 318418.0  215.5 42055.0 101120.0
Solaris  3560.3 2616.8 2892.9  2529.8 3428.9  2047.2 12327.0 17840.0
Up’n’Down  125754.6 82555.0 401884.3  324505.2 469548.3   707.2 11693.0 16080.0
Video Pinball  533936.5 19123.1 565163.2  243320.1 922518.0   20452.0 17668.0 32420.0
Yars’ Revenge  102557.0 61575.7 148594.8  109980.9 498947.1   1476.9 54577.0 83523.0
Table 3: Scores obtained by evaluating the best checkpoint for 200 episodes using the no-op starts regime.
Figure 6: The human-relative score of Ape-X DQfD (deeper) using the no-ops starts regime. The score is computed as .
Game  Rainbow DQfD Ape-X DQN  Ape-X DQfD Ape-X DQfD (deeper)  Random Avg. Human Expert
Alien  6022.9 17731.5   1025.5 6983.4  6371.3
Amidar  202.8 1047.3  310.5 1177.5  1540.4
Assault  14491.7 24404.6  23384.3 34716.5   628.9
Asterix  280114.0 283179.5  327929.0 297533.8  7536.0
Asteroids  2249.4 117303.4   95066.6 95170.9  36517.3
Atlantis  814684.0 918714.5  912443.0 1020311.0   26575.0
Bank Heist  826.0 1200.8  1695.9 2020.5   644.5
Battle Zone  52040.0 92275.0   42150.0 74410.0  33030.0
Beam Rider  21768.5 72233.7  46454.5 82997.1   14961.0
Bowling  39.4 30.2  178.3 174.4  146.5
Boxing  54.9 80.9   64.5 69.7  9.6
Breakout  379.5 756.5   145.1 365.5  27.9
Chopper Command  10916.0 576601.5  90152.5 681202.5   8930.0
Crazy Climber  143962.0 263953.5   141468.0 196633.5  32667.0
Defender  47671.3 399865.3   37771.8 123734.8  14296.0
Demon Attack  109670.7 133002.1  97458.8 142189.0   3442.8
Double Dunk  -0.6 22.3   20.5 21.8  -14.4
Enduro  2061.1 2042.4  1538.3 1754.9  740.2
Fishing Derby  22.6 22.4  26.3 24.0  5.1
Freeway  29.1 29.0  23.8 26.8  25.6
Gopher  72595.7 121168.2   115654.7 115392.1  2311.0
Gravitar  567.5 662.0  972.0 1021.8  3116.0
Hero  50496.8 26345.3  104942.1 107144.0   25839.4
Ice Hockey  -0.7 24.0   3.3 18.4  0.5
James Bond  18142.3 18992.3   12041.0 15010.0  368.5
Kangaroo  10841.0 577.5  25953.5 28616.0   2739.0
Krull  6715.5 8592.0  111496.1 122870.1   2109.1
Kung Fu Master  28999.8 72068.0  50421.5 102258.0   20786.8
Montezuma’s Revenge  154.0 1079.0  22781.0 22730.5  4182.0
Ms. Pacman  2570.2 6135.4  1880.8 4007.4  15375.0
Name This Game  11686.5 23829.9  22874.6 29416.0   6796.0
Pitfall!  -37.6 -273.3  3367.5 3208.7  5998.9
Pong  19.0 18.7  14.0 18.6  15.5
Private Eye  1704.4 864.7  61895.1 54976.0  64169.1
Q*bert  18397.6 380152.1   41419.6 51159.3  12085.0
Riverraid  15608.1 49982.8   18720.1 42288.9  14382.2
Road Runner  54261.0 127111.5  486082.0 507490.0   6878.0
Seaquest  19176.0 377179.8   15526.1 269480.0  40425.8
Solaris  2860.7 3115.9  2235.6 1835.8  11032.6
Up’n’Down  92640.6 347912.2   200709.3 298361.8  9896.1
Video Pinball  506817.2 873988.5   194845.0 832691.1  15641.1
Yars’ Revenge  93007.9 131701.1  82521.8 466181.8   47135.2
Table 4: Scores obtained by evaluating the best checkpoint for 200 episodes using the human starts regime.
Figure 7: The human-relative score of Ape-X DQfD (deeper) using the human starts regime. The score is computed as .

Appendix E Experimental setup & hyper parameters

Game  Min score Max score Number of transitions Number of episodes
Alien  9690 29160 19133 5
Amidar  1353 2341 16790 5
Assault  1168 2274 13224 5
Asterix  4500 18100 9525 5
Asteroids  14170 18100 22801 5
Atlantis  10300 22400 17516 12
Bank Heist  900 7465 32389 7
Battle Zone  35000 60000 9075 5
Beam Rider  12594 19844 38665 4
Bowling  89 149 9991 5
Boxing  0 15 8438 5
Breakout  17 79 10475 9
Chopper Command  4700 11300 7710 5
Crazy Climber  30600 61600 18937 5
Defender  5150 18700 6421 5
Demon Attack  1800 6190 17409 5
Double Dunk  -22 -14 11855 5
Enduro  383 803 42058 5
Fishing Derby  -10 20 6388 4
Freeway  30 32 10239 5
Gopher  2500 22520 38632 5
Gravitar  2950 13400 15377 5
Hero  35155 99320 32907 5
Ice Hockey  -4 1 17585 5
James Bond  400 650 9050 5
Kangaroo  12400 36300 20984 5
Krull  8040 13730 32581 5
Kung Fu Master  8300 25920 12989 5
Montezuma’s Revenge  32300 34900 17949 5
Ms Pacman  31781 55021 21896 3
Name This Game  11350 19380 43571 5
Pitfall  3662 47821 35347 5
Pong  -12 0 17719 3
Private Eye  70375 74456 10899 5
Q-Bert  80700 99450 75472 5
River Raid  17240 39710 46233 5
Road Runner  8400 20200 5574 5
Seaquest  56510 101120 57453 7
Solaris  2840 17840 28552 6
Up N Down  6580 16080 10421 4
Video Pinball  8409 32420 10051 5
Yars’ Revenge  48361 83523 21334 4
Table 5: The table shows the performance of our expert player and the amount of available demonstrations per game. Note that the total number of episodes/trajectories is very low.
Figure 8: The two network architectures that we used. The upper one is the standard dueling architecture of Wang et al. [2016] and the lower one is a slightly wider and deeper version.
Parameter  Comment Value
Learner configuration
Batch size  256
Agent transitions per batch  192