Transferring knowledge across domains is one of the most fundamental problems in machine learning, but doing so effectively in the context of reinforcement learning remains largely an open problem. Current methods make strong assumptions on the specifics of the task, often lack principled objectives, and -- crucially -- modify individual policies, which might be sub-optimal when the domains differ due to a drift in the state space, i.e., it is intrinsic to the environment and therefore affects every agent interacting with it. To address these drawbacks, we propose TvD: transfer via distribution matching, a framework to transfer knowledge across interactive domains. We approach the problem from a data-centric perspective, characterizing the discrepancy in environments by means of (potentially complex) transformation between their state spaces, and thus posing the problem of transfer as learning to undo this transformation. To accomplish this, we introduce a novel optimization objective based on an optimal transport distance between two distributions over trajectories -- those generated by an already-learned policy in the source domain and a learnable pushforward policy in the target domain. We show this objective leads to a policy update scheme reminiscent of imitation learning, and derive an efficient algorithm to implement it. Our experiments in simple gridworlds show that this method yields successful transfer learning across a wide range of environment transformations.

How would an artificial intelligence (AI) decide what to do? One common approach in AI research is called "reinforcement learning". Reinforcement learning gives the software a "reward" defined in some way, and lets the software figure out how to maximise the reward. This approach has produced some excellent results, such as building software agents that defeat humans at games like chess and Go, or creating new designs for nuclear fusion reactors. However, we might want to hold off on making reinforcement learning agents too flexible and effective.

In this work, we consider one-shot imitation learning for object rearrangement tasks, where an AI agent needs to watch a single expert demonstration and learn to perform the same task in different environments. To achieve a strong generalization, the AI agent must infer the spatial goal specification for the task. However, there can be multiple goal specifications that fit the given demonstration. To address this, we propose a reward learning approach, Graph-based Equivalence Mappings (GEM), that can discover spatial goal representations that are aligned with the intended goal specification, enabling successful generalization in unseen environments. Specifically, GEM represents a spatial goal specification by a reward function conditioned on i) a graph indicating important spatial relationships between objects and ii) state equivalence mappings for each edge in the graph indicating invariant properties of the corresponding relationship. GEM combines inverse reinforcement learning and active reward learning to efficiently improve the reward function by utilizing the graph structure and domain randomization enabled by the equivalence mappings. We conducted experiments with simulated oracles and with human subjects. The results show that GEM can drastically improve the generalizability of the learned goal representations over strong baselines.

Graph Neural Networks (GNNs) have drawn significant attentions over the years and been broadly applied to essential applications requiring solid robustness or vigorous security standards, such as product recommendation and user behavior modeling. Under these scenarios, exploiting GNN's vulnerabilities and further downgrading its performance become extremely incentive for adversaries. Previous attackers mainly focus on structural perturbations or node injections to the existing graphs, guided by gradients from the surrogate models. Although they deliver promising results, several limitations still exist. For the structural perturbation attack, to launch a proposed attack, adversaries need to manipulate the existing graph topology, which is impractical in most circumstances. Whereas for the node injection attack, though being more practical, current approaches require training surrogate models to simulate a white-box setting, which results in significant performance downgrade when the surrogate architecture diverges from the actual victim model. To bridge these gaps, in this paper, we study the problem of black-box node injection attack, without training a potentially misleading surrogate model. Specifically, we model the node injection attack as a Markov decision process and propose Gradient-free Graph Advantage Actor Critic, namely G2A2C, a reinforcement learning framework in the fashion of advantage actor critic. By directly querying the victim model, G2A2C learns to inject highly malicious nodes with extremely limited attacking budgets, while maintaining a similar node feature distribution. Through our comprehensive experiments over eight acknowledged benchmark datasets with different characteristics, we demonstrate the superior performance of our proposed G2A2C over the existing state-of-the-art attackers. Source code is publicly available at: https://github.com/jumxglhf/G2A2C}.

No human drives a car in a vacuum; she/he must negotiate with other road users to achieve their goals in social traffic scenes. A rational human driver can interact with other road users in a socially-compatible way through implicit communications to complete their driving tasks smoothly in interaction-intensive, safety-critical environments. This paper aims to review the existing approaches and theories to help understand and rethink the interactions among human drivers toward social autonomous driving. We take this survey to seek the answers to a series of fundamental questions: 1) What is social interaction in road traffic scenes? 2) How to measure and evaluate social interaction? 3) How to model and reveal the process of social interaction? 4) How do human drivers reach an implicit agreement and negotiate smoothly in social interaction? This paper reviews various approaches to modeling and learning the social interactions between human drivers, ranging from optimization theory and graphical models to social force theory and behavioral & cognitive science. We also highlight some new directions, critical challenges, and opening questions for future research.

The predictive information, the mutual information between the past and future, has been shown to be a useful representation learning auxiliary loss for training reinforcement learning agents, as the ability to model what will happen next is critical to success on many control tasks. While existing studies are largely restricted to training specialist agents on single-task settings in simulation, in this work, we study modeling the predictive information for robotic agents and its importance for general-purpose agents that are trained to master a large repertoire of diverse skills from large amounts of data. Specifically, we introduce Predictive Information QT-Opt (PI-QT-Opt), a QT-Opt agent augmented with an auxiliary loss that learns representations of the predictive information to solve up to 297 vision-based robot manipulation tasks in simulation and the real world with a single set of parameters. We demonstrate that modeling the predictive information significantly improves success rates on the training tasks and leads to better zero-shot transfer to unseen novel tasks. Finally, we evaluate PI-QT-Opt on real robots, achieving substantial and consistent improvement over QT-Opt in multiple experimental settings of varying environments, skills, and multi-task configurations.

Multi-agent systems in which secondary agents with conflicting agendas also alter their methods need opponent modeling. In this study, we simulate the main agent's and secondary agents' tactics using Double Deep Q-Networks (DDQN) with a prioritized experience replay mechanism. Then, under the opponent modeling setup, a Mixture-of-Experts architecture is used to identify various opponent strategy patterns. Finally, we analyze our models in two environments with several agents. The findings indicate that the Mixture-of-Experts model, which is based on opponent modeling, performs better than DDQN.

Increasing traffic demands, higher levels of automation, and communication enhancements provide novel design opportunities for future air traffic controllers (ATCs). This article presents a novel deep reinforcement learning (DRL) controller to aid conflict resolution for autonomous free flight. Although DRL has achieved important advancements in this field, the existing works pay little attention to the explainability and safety issues related to DRL controllers, particularly the safety under adversarial attacks. To address those two issues, we design a fully explainable DRL framework wherein we: 1) decompose the coupled Q value learning model into a safety-awareness and efficiency (reach the target) one; and 2) use information from surrounding intruders as inputs, eliminating the needs of central controllers. In our simulated experiments, we show that by decoupling the safety-awareness and efficiency, we can exceed performance on free flight control tasks while dramatically improving explainability on practical. In addition, the safety Q learning module provides rich information about the safety situation of environments. To study the safety under adversarial attacks, we additionally propose an adversarial attack strategy that can impose both safety-oriented and efficiency-oriented attacks. The adversarial aims to minimize safety/efficiency by only attacking the agent at a few time steps. In the experiments, our attack strategy increases as many collisions as the uniform attack (i.e., attacking at every time step) by only attacking the agent four times less often, which provide insights into the capabilities and restrictions of the DRL in future ATC designs. The source code is publicly available at https://github.com/WLeiiiii/Gym-ATC-Attack-Project.

Recent years have witnessed a large amount of decentralized data in various (edge) devices of end-users, while the decentralized data aggregation remains complicated for machine learning jobs because of regulations and laws. As a practical approach to handling decentralized data, Federated Learning (FL) enables collaborative global machine learning model training without sharing sensitive raw data. The servers schedule devices to jobs within the training process of FL. In contrast, device scheduling with multiple jobs in FL remains a critical and open problem. In this paper, we propose a novel multi-job FL framework, which enables the training process of multiple jobs in parallel. The multi-job FL framework is composed of a system model and a scheduling method. The system model enables a parallel training process of multiple jobs, with a cost model based on the data fairness and the training time of diverse devices during the parallel training process. We propose a novel intelligent scheduling approach based on multiple scheduling methods, including an original reinforcement learning-based scheduling method and an original Bayesian optimization-based scheduling method, which corresponds to a small cost while scheduling devices to multiple jobs. We conduct extensive experimentation with diverse jobs and datasets. The experimental results reveal that our proposed approaches significantly outperform baseline approaches in terms of training time (up to 12.73 times faster) and accuracy (up to 46.4% higher).

Information-directed sampling (IDS) has revealed its potential as a data-efficient algorithm for reinforcement learning (RL). However, theoretical understanding of IDS for Markov Decision Processes (MDPs) is still limited. We develop novel information-theoretic tools to bound the information ratio and cumulative information gain about the learning target. Our theoretical results shed light on the importance of choosing the learning target such that the practitioners can balance the computation and regret bounds. As a consequence, we derive prior-free Bayesian regret bounds for vanilla-IDS which learns the whole environment under tabular finite-horizon MDPs. In addition, we propose a computationally-efficient regularized-IDS that maximizes an additive form rather than the ratio form and show that it enjoys the same regret bound as vanilla-IDS. With the aid of rate-distortion theory, we improve the regret bound by learning a surrogate, less informative environment. Furthermore, we extend our analysis to linear MDPs and prove similar regret bounds for Thompson sampling as a by-product.