Search or filter publications

Filter by type:

Filter by publication type

Filter by year:



  • Showing results for:
  • Reset all filters

Search results

  • Conference paper
    Tavakoli A, Pardo F, Kormushev P, 2018,

    Action branching architectures for deep reinforcement learning

    , AAAI 2018, Publisher: AAAI

    Discrete-action algorithms have been central to numerousrecent successes of deep reinforcement learning. However,applying these algorithms to high-dimensional action tasksrequires tackling the combinatorial increase of the numberof possible actions with the number of action dimensions.This problem is further exacerbated for continuous-actiontasks that require fine control of actions via discretization.In this paper, we propose a novel neural architecture fea-turing a shared decision module followed by several net-workbranches, one for each action dimension. This approachachieves a linear increase of the number of network outputswith the number of degrees of freedom by allowing a level ofindependence for each individual action dimension. To illus-trate the approach, we present a novel agent, called Branch-ing Dueling Q-Network (BDQ), as a branching variant ofthe Dueling Double Deep Q-Network (Dueling DDQN). Weevaluate the performance of our agent on a set of challeng-ing continuous control tasks. The empirical results show thatthe proposed agent scales gracefully to environments with in-creasing action dimensionality and indicate the significanceof the shared decision module in coordination of the dis-tributed action branches. Furthermore, we show that the pro-posed agent performs competitively against a state-of-the-art continuous control algorithm, Deep Deterministic PolicyGradient (DDPG).

  • Journal article
    Chamberlain B, Levy-Kramer J, Humby C, Deisenroth MPet al., 2018,

    Real-time community detection in full social networks on a laptop

    , PLoS ONE, Vol: 13, ISSN: 1932-6203

    For a broad range of research and practical applications it is important to understand the allegiances, communities and structure of key players in society. One promising direction towards extracting this information is to exploit the rich relational data in digital social networks (the social graph). As global social networks (e.g., Facebook and Twitter) are very large, most approaches make use of distributed computing systems for this purpose. Distributing graph processing requires solving many difficult engineering problems, which has lead some researchers to look at single-machine solutions that are faster and easier to maintain. In this article, we present an approach for analyzing full social networks on a standard laptop, allowing for interactive exploration of the communities in the locality of a set of user specified query vertices. The key idea is that the aggregate actions of large numbers of users can be compressed into a data structure that encapsulates the edge weights between vertices in a derived graph. Local communities can be constructed by selecting vertices that are connected to the query vertices with high edge weights in the derived graph. This compression is robust to noise and allows for interactive queries of local communities in real-time, which we define to be less than the average human reaction time of 0.25s. We achieve single-machine real-time performance by compressing the neighborhood of each vertex using minhash signatures and facilitate rapid queries through Locality Sensitive Hashing. These techniques reduce query times from hours using industrial desktop machines operating on the full graph to milliseconds on standard laptops. Our method allows exploration of strongly associated regions (i.e., communities) of large graphs in real-time on a laptop. It has been deployed in software that is actively used by social network analysts and offers another channel for media owners to monetize their data, helping them to continue to provide

  • Conference paper
    Kanajar P, Caldwell DG, Kormushev P, 2017,

    Climbing over large obstacles with a humanoid robot via multi-contact motion planning

    , IEEE RO-MAN 2017: 26th IEEE International Symposium on Robot and Human Interactive Communication, Publisher: IEEE, Pages: 1202-1209

    Incremental progress in humanoid robot locomotion over the years has achieved important capabilities such as navigation over flat or uneven terrain, stepping over small obstacles and climbing stairs. However, the locomotion research has mostly been limited to using only bipedal gait and only foot contacts with the environment, using the upper body for balancing without considering additional external contacts. As a result, challenging locomotion tasks like climbing over large obstacles relative to the size of the robot have remained unsolved. In this paper, we address this class of open problems with an approach based on multi-body contact motion planning guided through physical human demonstrations. Our goal is to make the humanoid locomotion problem more tractable by taking advantage of objects in the surrounding environment instead of avoiding them. We propose a multi-contact motion planning algorithm for humanoid robot locomotion which exploits the whole-body motion and multi-body contacts including both the upper and lower body limbs. The proposed motion planning algorithm is applied to a challenging task of climbing over a large obstacle. We demonstrate successful execution of the climbing task in simulation using our multi-contact motion planning algorithm initialized via a transfer from real-world human demonstrations of the task and further optimized.

  • Conference paper
    Salimbeni H, Deisenroth M, 2017,

    Doubly stochastic variational inference for deep Gaussian processes

    , NIPS 2017, Publisher: Advances in Neural Information Processing Systems (NIPS), Pages: 4589-4600, ISSN: 1049-5258

    Gaussian processes (GPs) are a good choice for function approximation as theyare flexible, robust to over-fitting, and provide well-calibrated predictiveuncertainty. Deep Gaussian processes (DGPs) are multi-layer generalisations ofGPs, but inference in these models has proved challenging. Existing approachesto inference in DGP models assume approximate posteriors that forceindependence between the layers, and do not work well in practice. We present adoubly stochastic variational inference algorithm, which does not forceindependence between layers. With our method of inference we demonstrate that aDGP model can be used effectively on data ranging in size from hundreds to abillion points. We provide strong empirical evidence that our inference schemefor DGPs works well in practice in both classification and regression.

  • Conference paper
    Rakicevic N, Kormushev P, 2017,

    Efficient Robot Task Learning and Transfer via Informed Search in Movement Parameter Space

    , Workshop on Acting and Interacting in the Real World: Challenges in Robot Learning, 31st Conference on Neural Information Processing Systems (NIPS 2017)
  • Conference paper
    Tavakoli A, Pardo F, Kormushev P, 2017,

    Action Branching Architectures for Deep Reinforcement Learning

    , Deep Reinforcement Learning Symposium, 31st Conference on Neural Information Processing Systems (NIPS 2017)
  • Conference paper
    Eleftheriadis S, Nicholson TFW, Deisenroth MP, Hensman Jet al., 2017,

    Identification of Gaussian Process State Space Models

    , Advances in Neural Information Processing Systems (NIPS) 2017, Publisher: Neural Information Processing Systems Foundation, Inc., Pages: 5310-5320, ISSN: 1049-5258

    The Gaussian process state space model (GPSSM) is a non-linear dynamicalsystem, where unknown transition and/or measurement mappings are described byGPs. Most research in GPSSMs has focussed on the state estimation problem.However, the key challenge in GPSSMs has not been satisfactorily addressed yet:system identification. To address this challenge, we impose a structuredGaussian variational posterior distribution over the latent states, which isparameterised by a recognition model in the form of a bi-directional recurrentneural network. Inference with this structure allows us to recover a posteriorsmoothed over the entire sequence(s) of data. We provide a practical algorithmfor efficiently computing a lower bound on the marginal likelihood using thereparameterisation trick. This additionally allows arbitrary kernels to be usedwithin the GPSSM. We demonstrate that we can efficiently generate plausiblefuture trajectories of the system we seek to model with the GPSSM, requiringonly a small number of interactions with the true system.

  • Journal article
    Huang R, Lattimore T, Gyorgy A, Szepesvari Cet al., 2017,

    Following the Leader and Fast Rates in Online Linear Prediction: Curved Constraint Sets and Other Regularities

    , Journal of Machine Learning Research, Vol: 18(145), Pages: 1-31, ISSN: 1532-4435

    Follow the leader (FTL) is a simple online learning algorithm that is known to perform well whenthe loss functions are convex and positively curved. In this paper we ask whether there are othersettings when FTL achieves low regret. In particular, we study the fundamental problem of linearprediction over a convex, compact domain with non-empty interior. Amongst other results, weprove that the curvature of the boundary of the domain can act as if the losses were curved: In thiscase, we prove that as long as the mean of the loss vectors have positive lengths bounded away fromzero, FTL enjoys logarithmic regret, while for polytope domains and stochastic data it enjoys finiteexpected regret. The former result is also extended to strongly convex domains by establishing anequivalence between the strong convexity of sets and the minimum curvature of their boundary,which may be of independent interest. Building on a previously known meta-algorithm, we alsoget an algorithm that simultaneously enjoys the worst-case guarantees and the smaller regret ofFTL when the data is ‘easy’. Finally, we show that such guarantees are achievable directly (e.g.,by the follow the regularized leader algorithm or by a shrinkage-based variant of FTL) when theconstraint set is an ellipsoid.

  • Conference paper
    Kamthe S, Deisenroth MP, 2018,

    Data-Efficient Reinforcement Learning with Probabilistic Model Predictive Control

    , International Conference on Artificial Intelligence and Statistics

    Trial-and-error based reinforcement learning (RL) has seen rapid advancementsin recent times, especially with the advent of deep neural networks. However,the majority of autonomous RL algorithms either rely on engineered features ora large number of interactions with the environment. Such a large number ofinteractions may be impractical in many real-world applications. For example,robots are subject to wear and tear and, hence, millions of interactions maychange or damage the system. Moreover, practical systems have limitations inthe form of the maximum torque that can be safely applied. To reduce the numberof system interactions while naturally handling constraints, we propose amodel-based RL framework based on Model Predictive Control (MPC). Inparticular, we propose to learn a probabilistic transition model using GaussianProcesses (GPs) to incorporate model uncertainties into long-term predictions,thereby, reducing the impact of model errors. We then use MPC to find a controlsequence that minimises the expected long-term cost. We provide theoreticalguarantees for the first-order optimality in the GP-based transition modelswith deterministic approximate inference for long-term planning. The proposedframework demonstrates superior data efficiency and learning rates compared tothe current state of the art.

  • Journal article
    Arulkumaran K, Deisenroth MP, Brundage M, Bharath AAet al., 2017,

    A brief survey of deep reinforcement learning

    , IEEE Signal Processing Magazine, Vol: 34, Pages: 26-38, ISSN: 1053-5888

    Deep reinforcement learning (DRL) is poised to revolutionize the field of artificial intelligence (AI) and represents a step toward building autonomous systems with a higherlevel understanding of the visual world. Currently, deep learning is enabling reinforcement learning (RL) to scale to problems that were previously intractable, such as learning to play video games directly from pixels. DRL algorithms are also applied to robotics, allowing control policies for robots to be learned directly from camera inputs in the real world. In this survey, we begin with an introduction to the general field of RL, then progress to the main streams of value-based and policy-based methods. Our survey will cover central algorithms in deep RL, including the deep Q-network (DQN), trust region policy optimization (TRPO), and asynchronous advantage actor critic. In parallel, we highlight the unique advantages of deep neural networks, focusing on visual understanding via RL. To conclude, we describe several current areas of research within the field.

This data is extracted from the Web of Science and reproduced under a licence from Thomson Reuters. You may not copy or re-distribute this data in whole or in part without the written consent of the Science business of Thomson Reuters.

Request URL: Request URI: /respub/WEB-INF/jsp/search-t4-html.jsp Query String: id=954&limit=10&page=4&respub-action=search.html Current Millis: 1603784814840 Current Time: Tue Oct 27 07:46:54 GMT 2020