Large-batch training has become a commonly used technique when training neural networks with a large number of GPU/TPU processors. As batch size increases, stochastic optimizers tend to converge to sharp local minima, leading to degraded test performance. Current methods usually use extensive data augmentation to increase the batch size, but we found the performance gain with data augmentation decreases as batch size increases, and data augmentation will become insufficient after certain point. In this paper, we propose to use adversarial learning to increase the batch size in large-batch training. Despite being a natural choice for smoothing the decision surface and biasing towards a flat region, adversarial learning has not been successfully applied in large-batch training since it requires at least two sequential gradient computations at each step, which will at least double the running time compared with vanilla training even with a large number of processors. To overcome this issue, we propose a novel Concurrent Adversarial Learning (ConAdv) method that decouple the sequential gradient computations in adversarial learning by utilizing staled parameters. Experimental results demonstrate that ConAdv can successfully increase the batch size on both ResNet-50 and EfficientNet training on ImageNet while maintaining high accuracy. In particular, we show ConAdv along can achieve 75.3\% top-1 accuracy on ImageNet ResNet-50 training with 96K batch size, and the accuracy can be further improved to 76.2\% when combining ConAdv with data augmentation. This is the first work successfully scales ResNet-50 training batch size to 96K.

Maximising a cumulative reward function that is Markov and stationary, i.e., defined over state-action pairs and independent of time, is sufficient to capture many kinds of goals in a Markov Decision Process (MDP) based on the Reinforcement Learning (RL) problem formulation. However, not all goals can be captured in this manner. Specifically, it is easy to see that Convex MDPs in which goals are expressed as convex functions of stationary distributions cannot, in general, be formulated in this manner. In this paper, we reformulate the convex MDP problem as a min-max game between the policy and cost (negative reward) players using Fenchel duality and propose a meta-algorithm for solving it. We show that the average of the policies produced by an RL agent that maximizes the non-stationary reward produced by the cost player converges to an optimal solution to the convex MDP. Finally, we show that the meta-algorithm unifies several disparate branches of reinforcement learning algorithms in the literature, such as apprenticeship learning, variational intrinsic control, constrained MDPs, and pure exploration into a single framework.

Goal-conditioned hierarchical reinforcement learning (HRL) serves as a successful approach to solving complex and temporally extended tasks. Recently, its success has been extended to more general settings by concurrently learning hierarchical policies and subgoal representations. However, online subgoal representation learning exacerbates the non-stationary issue of HRL and introduces challenges for exploration in high-level policy learning. In this paper, we propose a state-specific regularization that stabilizes subgoal embeddings in well-explored areas while allowing representation updates in less explored state regions. Benefiting from this stable representation, we design measures of novelty and potential for subgoals, and develop an efficient hierarchical exploration strategy that actively seeks out new promising subgoals and states. Experimental results show that our method significantly outperforms state-of-the-art baselines in continuous control tasks with sparse rewards and further demonstrate the stability and efficiency of the subgoal representation learning of this work, which promotes superior policy learning.

In this work we study how to learn good algorithms for selecting reasoning steps in theorem proving. We explore this in the connection tableau calculus implemented by leanCoP where the partial tableau provides a clean and compact notion of a state to which a limited number of inferences can be applied. We start by incorporating a state-of-the-art learning algorithm -- a graph neural network (GNN) -- into the plCoP theorem prover. Then we use it to observe the system's behaviour in a reinforcement learning setting, i.e., when learning inference guidance from successful Monte-Carlo tree searches on many problems. Despite its better pattern matching capability, the GNN initially performs worse than a simpler previously used learning algorithm. We observe that the simpler algorithm is less confident, i.e., its recommendations have higher entropy. This leads us to explore how the entropy of the inference selection implemented via the neural network influences the proof search. This is related to research in human decision-making under uncertainty, and in particular the probability matching theory. Our main result shows that a proper entropy regularisation, i.e., training the GNN not to be overconfident, greatly improves plCoP's performance on a large mathematical corpus.

Despite the success of reinforcement learning methods, they have yet to have their breakthrough moment when applied to a broad range of robotic manipulation tasks. This is partly due to the fact that reinforcement learning algorithms are notoriously difficult and time consuming to train, which is exacerbated when training from images rather than full-state inputs. As humans perform manipulation tasks, our eyes closely monitor every step of the process with our gaze focusing sequentially on the objects being manipulated. With this in mind, we present our Attention-driven Robotic Manipulation (ARM) algorithm, which is a general manipulation algorithm that can be applied to a range of sparse-rewarded tasks, given only a small number of demonstrations. ARM splits the complex task of manipulation into a 3 stage pipeline: (1) a Q-attention agent extracts interesting pixel locations from RGB and point cloud inputs, (2) a next-best pose agent that accepts crops from the Q-attention agent and outputs poses, and (3) a control agent that takes the goal pose and outputs joint actions. We show that current learning algorithms fail on a range of RLBench tasks, whilst ARM is successful.

Value factorisation proves to be a very useful technique in multi-agent reinforcement learning (MARL), but the underlying mechanism is not yet fully understood. This paper explores a theoretic basis for value factorisation. We generalise the Shapley value in the coalitional game theory to a Markov convex game (MCG) and use it to guide value factorisation in MARL. We show that the generalised Shapley value possesses several features such as (1) accurate estimation of the maximum global value, (2) fairness in the factorisation of the global value, and (3) being sensitive to dummy agents. The proposed theory yields a new learning algorithm called Sharpley Q-learning (SHAQ), which inherits the important merits of ordinary Q-learning but extends it to MARL. In comparison with prior-arts, SHAQ has a much weaker assumption (MCG) that is more compatible with real-world problems, but has superior explainability and performance in many cases. We demonstrated SHAQ and verified the theoretic claims on Predator-Prey and StarCraft Multi-Agent Challenge (SMAC).

Reparameterization (RP) and likelihood ratio (LR) gradient estimators are used to estimate gradients of expectations throughout machine learning and reinforcement learning; however, they are usually explained as simple mathematical tricks, with no insight into their nature. We use a first principles approach to explain that LR and RP are alternative methods of keeping track of the movement of probability mass, and the two are connected via the divergence theorem. Moreover, we show that the space of all possible estimators combining LR and RP can be completely parameterized by a flow field $u(x)$ and an importance sampling distribution $q(x)$. We prove that there cannot exist a single-sample estimator of this type outside our characterized space, thus, clarifying where we should be searching for better Monte Carlo gradient estimators.

We chose Ray because we needed to train many reinforcement learning agents simultaneously. It was important to us to deliver results quickly to people using Pathmind, which simulation modelers use to apply reinforcement learning to industrial operations and supply chains. We use Ray, RLlib and Ray Serve so that businesses can quickly train RL on the cloud and feed RL-based decisions to their operations for use cases that range from optimal warehouse picking to routing autonomous guided vehicles.

Amazon today launched SageMaker Reinforcement Learning (RL) Kubeflow Components, a toolkit supporting the company's AWS RoboMaker service for orchestrating robotics workflows. Amazon says that the goal is to make it faster to experiment and manage robotics workloads from perception to controls and optimization, and to create end-to-end solutions without having to rebuild them each time. Robots are being used more widely for purposes that are increasing in sophistication, like assembly, picking and packing, last-mile delivery, environmental monitoring, search and rescue, and assisted surgery. In China, Oxford Economics anticipates 12.5 million manufacturing jobs will become automated, while in the U.S., McKinsey projects that machines will take upwards of 30% of such jobs. As for reinforcement learning, it's an emerging AI technique that can help develop solutions for the kinds of problems that are increasingly cropping up in robotics.

Deep reinforcement learning (DRL) has been demonstrated to provide promising results in several challenging decision making and control tasks. However, the required inference costs of deep neural networks (DNNs) could prevent DRL from being applied to mobile robots which cannot afford high energy-consuming computations. To enable DRL methods to be affordable in such energy-limited platforms, we propose an asymmetric architecture that reduces the overall inference costs via switching between a computationally expensive policy and an economic one. The experimental results evaluated on a number of representative benchmark suites for robotic control tasks demonstrate that our method is able to reduce the inference costs while retaining the agent's overall performance.