While multi-agent trust region algorithms have achieved great success empirically in solving coordination tasks, most of them, however, suffer from a non-stationarity problem since agents update their policies simultaneously. In contrast, a sequential scheme that updates policies agent-by-agent provides another perspective and shows strong performance. However, sample inefficiency and lack of monotonic improvement guarantees for each agent are still the two significant challenges for the sequential scheme. In this paper, we propose the \textbf{A}gent-by-\textbf{a}gent \textbf{P}olicy \textbf{O}ptimization (A2PO) algorithm to improve the sample efficiency and retain the guarantees of monotonic improvement for each agent during training. We justify the tightness of the monotonic improvement bound compared with other trust region algorithms. From the perspective of sequentially updating agents, we further consider the effect of agent updating order and extend the theory of non-stationarity into the sequential update scheme. To evaluate A2PO, we conduct a comprehensive empirical study on four benchmarks: StarCraftII, Multi-agent MuJoCo, Multi-agent Particle Environment, and Google Research Football full game scenarios. A2PO consistently outperforms strong baselines.
Real-world data usually exhibits a long-tailed distribution,with a few frequent labels and a lot of few-shot labels. The study of institution name normalization is a perfect application case showing this phenomenon. There are many institutions worldwide with enormous variations of their names in the publicly available literature. In this work, we first collect a large-scale institution name normalization dataset LoT-insts1, which contains over 25k classes that exhibit a naturally long-tailed distribution. In order to isolate the few-shot and zero-shot learning scenarios from the massive many-shot classes, we construct our test set from four different subsets: many-, medium-, and few-shot sets, as well as a zero-shot open set. We also replicate several important baseline methods on our data, covering a wide range from search-based methods to neural network methods that use the pretrained BERT model. Further, we propose our specially pretrained, BERT-based model that shows better out-of-distribution generalization on few-shot and zero-shot test sets. Compared to other datasets focusing on the long-tailed phenomenon, our dataset has one order of magnitude more training data than the largest existing long-tailed datasets and is naturally long-tailed rather than manually synthesized. We believe it provides an important and different scenario to study this problem. To our best knowledge, this is the first natural language dataset that focuses on long-tailed and open-set classification problems.
Visual imitation learning enables reinforcement learning agents to learn to behave from expert visual demonstrations such as videos or image sequences, without explicit, well-defined rewards. Previous research either adopted supervised learning techniques or induce simple and coarse scalar rewards from pixels, neglecting the dense information contained in the image demonstrations. In this work, we propose to measure the expertise of various local regions of image samples, or called \textit{patches}, and recover multi-dimensional \textit{patch rewards} accordingly. Patch reward is a more precise rewarding characterization that serves as a fine-grained expertise measurement and visual explainability tool. Specifically, we present Adversarial Imitation Learning with Patch Rewards (PatchAIL), which employs a patch-based discriminator to measure the expertise of different local parts from given images and provide patch rewards. The patch-based knowledge is also used to regularize the aggregated reward and stabilize the training. We evaluate our method on DeepMind Control Suite and Atari tasks. The experiment results have demonstrated that PatchAIL outperforms baseline methods and provides valuable interpretations for visual demonstrations.
Graph Neural Networks (GNNs), originally proposed for node classification, have also motivated many recent works on edge prediction (a.k.a., link prediction). However, existing methods lack elaborate design regarding the distinctions between two tasks that have been frequently overlooked: (i) edges only constitute the topology in the node classification task but can be used as both the topology and the supervisions (i.e., labels) in the edge prediction task; (ii) the node classification makes prediction over each individual node, while the edge prediction is determinated by each pair of nodes. To this end, we propose a novel edge prediction paradigm named Edge-aware Message PassIng neuRal nEtworks (EMPIRE). Concretely, we first introduce an edge splitting technique to specify use of each edge where each edge is solely used as either the topology or the supervision (named as topology edge or supervision edge). We then develop a new message passing mechanism that generates the messages to source nodes (through topology edges) being aware of target nodes (through supervision edges). In order to emphasize the differences between pairs connected by supervision edges and pairs unconnected, we further weight the messages to highlight the relative ones that can reflect the differences. In addition, we design a novel negative node-pair sampling trick that efficiently samples 'hard' negative instances in the supervision instances, and can significantly improve the performance. Experimental results verify that the proposed method can significantly outperform existing state-of-the-art models regarding the edge prediction task on multiple homogeneous and heterogeneous graph datasets.
Our situated environment is full of uncertainty and highly dynamic, thus hindering the widespread adoption of machine-led Intelligent Decision-Making (IDM) in real world scenarios. This means IDM should have the capability of continuously learning new skills and efficiently generalizing across wider applications. IDM benefits from any new approaches and theoretical breakthroughs that exhibit Artificial General Intelligence (AGI) breaking the barriers between tasks and applications. Recent research has well-examined neural architecture, Transformer, as a backbone foundation model and its generalization to various tasks, including computer vision, natural language processing, and reinforcement learning. We therefore argue that a foundation decision model (FDM) can be established by formulating various decision-making tasks as a sequence decoding task using the Transformer architecture; this would be a promising solution to advance the applications of IDM in more complex real world tasks. In this paper, we elaborate on how a foundation decision model improves the efficiency and generalization of IDM. We also discuss potential applications of a FDM in multi-agent game AI, production scheduling, and robotics tasks. Finally, through a case study, we demonstrate our realization of the FDM, DigitalBrain (DB1) with 1.2 billion parameters, which achieves human-level performance over 453 tasks, including text generation, images caption, video games playing, robotic control, and traveling salesman problems. As a foundation decision model, DB1 would be a baby step towards more autonomous and efficient real world IDM applications.
Robots are traditionally bounded by a fixed embodiment during their operational lifetime, which limits their ability to adapt to their surroundings. Co-optimizing control and morphology of a robot, however, is often inefficient due to the complex interplay between the controller and morphology. In this paper, we propose a learning-based control method that can inherently take morphology into consideration such that once the control policy is trained in the simulator, it can be easily deployed to robots with different embodiments in the real world. In particular, we present the Embodiment-aware Transformer (EAT), an architecture that casts this control problem as conditional sequence modeling. EAT outputs the optimal actions by leveraging a causally masked Transformer. By conditioning an autoregressive model on the desired robot embodiment, past states, and actions, our EAT model can generate future actions that best fit the current robot embodiment. Experimental results show that EAT can outperform all other alternatives in embodiment-varying tasks, and succeed in an example of real-world evolution tasks: stepping down a stair through updating the morphology alone. We hope that EAT will inspire a new push toward real-world evolution across many domains, where algorithms like EAT can blaze a trail by bridging the field of evolutionary robotics and big data sequence modeling.
In reinforcement learning applications like robotics, agents usually need to deal with various input/output features when specified with different state/action spaces by their developers or physical restrictions. This indicates unnecessary re-training from scratch and considerable sample inefficiency, especially when agents follow similar solution steps to achieve tasks. In this paper, we aim to transfer similar high-level goal-transition knowledge to alleviate the challenge. Specifically, we propose PILoT, i.e., Planning Immediate Landmarks of Targets. PILoT utilizes the universal decoupled policy optimization to learn a goal-conditioned state planner; then, distills a goal-planner to plan immediate landmarks in a model-free style that can be shared among different agents. In our experiments, we show the power of PILoT on various transferring challenges, including few-shot transferring across action spaces and dynamics, from low-dimensional vector states to image inputs, from simple robot to complicated morphology; and we also illustrate a zero-shot transfer solution from a simple 2D navigation task to the harder Ant-Maze task.
Deep reinforcement learning has recently emerged as an appealing alternative for legged locomotion over multiple terrains by training a policy in physical simulation and then transferring it to the real world (i.e., sim-to-real transfer). Despite considerable progress, the capacity and scalability of traditional neural networks are still limited, which may hinder their applications in more complex environments. In contrast, the Transformer architecture has shown its superiority in a wide range of large-scale sequence modeling tasks, including natural language processing and decision-making problems. In this paper, we propose Terrain Transformer (TERT), a high-capacity Transformer model for quadrupedal locomotion control on various terrains. Furthermore, to better leverage Transformer in sim-to-real scenarios, we present a novel two-stage training framework consisting of an offline pretraining stage and an online correction stage, which can naturally integrate Transformer with privileged training. Extensive experiments in simulation demonstrate that TERT outperforms state-of-the-art baselines on different terrains in terms of return, energy consumption and control smoothness. In further real-world validation, TERT successfully traverses nine challenging terrains, including sand pit and stair down, which can not be accomplished by strong baselines.