Reward hacking--where RL agents exploit gaps in misspecified reward functions--has been widely observed, but not yet systematically studied. To understand how reward hacking arises, we construct four RL environments with misspecified rewards. We investigate reward hacking as a function of agent capabilities: model capacity, action space resolution, observation space noise, and training time. More capable agents often exploit reward misspecifications, achieving higher proxy reward and lower true reward than less capable agents. Moreover, we find instances of phase transitions: capability thresholds at which the agent's behavior qualitatively shifts, leading to a sharp decrease in the true reward. Such phase transitions pose challenges to monitoring the safety of ML systems. To address this, we propose an anomaly detection task for aberrant policies and offer several baseline detectors. As reinforcement learning agents are trained with better algorithms, more data, and larger policy models, they are at increased risk of overfitting their objectives (Russell, 2019). Reward hacking, or the gaming of misspecified reward functions by RL agents, has appeared in a variety of contexts, such as game playing (Ibarz et al., 2018), text summarization (Paulus et al., 2018), and autonomous driving (Knox et al., 2021). These examples show that better algorithms and models are not enough; for human-centered applications such as healthcare (Yu et al., 2019), economics (Trott et al., 2021) and robotics (Kober et al., 2013), RL algorithms must be safe and aligned with human objectives (Bommasani et al., 2021; Hubinger et al., 2019). Reward misspecifications occur because real-world tasks have numerous, often conflicting desiderata. In practice, reward designers resort to optimizing a proxy reward that is either more readily measured or more easily optimized than the true reward.

In this paper, we contribute a multi-faceted study into Pavlovian signalling -- a process by which learned, temporally extended predictions made by one agent inform decision-making by another agent. Signalling is intimately connected to time and timing. In service of generating and receiving signals, humans and other animals are known to represent time, determine time since past events, predict the time until a future stimulus, and both recognize and generate patterns that unfold in time. We investigate how different temporal processes impact coordination and signalling between learning agents by introducing a partially observable decision-making domain we call the Frost Hollow. In this domain, a prediction learning agent and a reinforcement learning agent are coupled into a two-part decision-making system that works to acquire sparse reward while avoiding time-conditional hazards. We evaluate two domain variations: machine agents interacting in a seven-state linear walk, and human-machine interaction in a virtual-reality environment. Our results showcase the speed of learning for Pavlovian signalling, the impact that different temporal representations do (and do not) have on agent-agent coordination, and how temporal aliasing impacts agent-agent and human-agent interactions differently. As a main contribution, we establish Pavlovian signalling as a natural bridge between fixed signalling paradigms and fully adaptive communication learning between two agents. We further show how to computationally build this adaptive signalling process out of a fixed signalling process, characterized by fast continual prediction learning and minimal constraints on the nature of the agent receiving signals. Our results therefore suggest an actionable, constructivist path towards communication learning between reinforcement learning agents.

We study what dataset assumption permits solving offline two-player zero-sum Markov game. In stark contrast to the offline single-agent Markov decision process, we show that the single strategy concentration assumption is insufficient for learning the Nash equilibrium (NE) strategy in offline two-player zero-sum Markov games. On the other hand, we propose a new assumption named unilateral concentration and design a pessimism-type algorithm that is provably efficient under this assumption. In addition, we show that the unilateral concentration assumption is necessary for learning an NE strategy. Furthermore, our algorithm can achieve minimax sample complexity without any modification for two widely studied settings: dataset with uniform concentration assumption and turn-based Markov game. Our work serves as an important initial step towards understanding offline multi-agent reinforcement learning.

Meanwhile, applications of RL have only begun to expand beyond these constrained game environments to more diverse and complex real-world environments such as chip design [86], chemical reaction optimization [133] and performing long-term recommendations [45]. To further progress towards these more complex real-world environments, greater alleviation of challenges currently facing RL (e.g., generalization, robustness, scalability, and safety) is needed [7, 27, 72, 108]. Moreover, we can expect that as the complexity of environments increases, the difficulty in alleviating these challenges will increase as well [27]. For the purpose of this paper, we broadly define known RL challenges as either an aptitude or alignment problem. Aptitude encompasses challenges concerned with being able to learn. Aptitude includes ideas such as robustness, the ability of RL to perform a task (e.g., asymptotic performance) and generalize within/between environments of similar complexity; scalability, the ability of RL to scale up to more complex environment; and aptness, the rate at which a RL algorithm can learn to solve a problem or achieve a desired performance level. Likewise, alignment encompasses challenges concerned with learning as intended [7, 27, 72]. The hypothetical paperclip agent [18] is a classic example of misalignment.

Reinforcement Learning is an aspect of Machine learning where an agent learns to behave in an environment, by performing certain actions and observing the rewards/results which it get from those actions. With the advancements in Robotics Arm Manipulation, Google Deep Mind beating a professional Alpha Go Player, and recently the OpenAI team beating a professional DOTA player, the field of reinforcement learning has really exploded in recent years. Applications of reinforcement learning were in the past limited by weak computer infrastructure. However, as Gerard Tesauro's backgamon AI superplayer developed in 1990's shows, progress did happen. That early progress is now rapidly changing with powerful new computational technologies opening the way to completely new inspiring applications. Training the models that control autonomous cars is an excellent example of a potential application of reinforcement learning.

Existing distributed cooperative multi-agent reinforcement learning (MARL) frameworks usually assume undirected coordination graphs and communication graphs while estimating a global reward via consensus algorithms for policy evaluation. Such a framework may induce expensive communication costs and exhibit poor scalability due to requirement of global consensus. In this work, we study MARLs with directed coordination graphs, and propose a distributed RL algorithm where the local policy evaluations are based on local value functions. The local value function of each agent is obtained by local communication with its neighbors through a directed learning-induced communication graph, without using any consensus algorithm. A zeroth-order optimization (ZOO) approach based on parameter perturbation is employed to achieve gradient estimation. By comparing with existing ZOO-based RL algorithms, we show that our proposed distributed RL algorithm guarantees high scalability. A distributed resource allocation example is shown to illustrate the effectiveness of our algorithm.

Federated learning (FL) is a training technique that enables client devices to jointly learn a shared model by aggregating locally-computed models without exposing their raw data. While most of the existing work focuses on improving the FL model accuracy, in this paper, we focus on the improving the training efficiency, which is often a hurdle for adopting FL in real-world applications. Specifically, we design an efficient FL framework which jointly optimizes model accuracy, processing latency and communication efficiency, all of which are primary design considerations for real implementation of FL. Inspired by the recent success of Multi-Agent Reinforcement Learning (MARL) in solving complex control problems, we present \textit{FedMarl}, an MARL-based FL framework which performs efficient run-time client selection. Experiments show that FedMarl can significantly improve model accuracy with much lower processing latency and communication cost.

The model-based reinforcement learning paradigm, which uses planning algorithms and neural network models, has recently achieved unprecedented results in diverse applications, leading to what is now known as deep reinforcement learning. These agents are quite complex and involve multiple components, factors that can create challenges for research. In this work, we propose a new modular software architecture suited for these types of agents, and a set of building blocks that can be easily reused and assembled to construct new model-based reinforcement learning agents. These building blocks include planning algorithms, policies, and loss functions. We illustrate the use of this architecture by combining several of these building blocks to implement and test agents that are optimized to three different test environments: Cartpole, Minigrid, and Tictactoe. One particular planning algorithm, made available in our implementation and not previously used in reinforcement learning, which we called averaged minimax, achieved good results in the three tested environments. Experiments performed with this architecture have shown that the best combination of planning algorithm, policy, and loss function is heavily problem dependent. This result provides evidence that the proposed architecture, which is modular and reusable, is useful for reinforcement learning researchers who want to study new environments and techniques.

Simultaneous segmentation and detection of liver tumors (hemangioma and hepatocellular carcinoma (HCC)) by using multi-modality non-contrast magnetic resonance imaging (NCMRI) are crucial for the clinical diagnosis. However, it is still a challenging task due to: (1) the HCC information on NCMRI is invisible or insufficient makes extraction of liver tumors feature difficult; (2) diverse imaging characteristics in multi-modality NCMRI causes feature fusion and selection difficult; (3) no specific information between hemangioma and HCC on NCMRI cause liver tumors detection difficult. In this study, we propose a united adversarial learning framework (UAL) for simultaneous liver tumors segmentation and detection using multi-modality NCMRI. The UAL first utilizes a multi-view aware encoder to extract multi-modality NCMRI information for liver tumor segmentation and detection. In this encoder, a novel edge dissimilarity feature pyramid module is designed to facilitate the complementary multi-modality feature extraction. Second, the newly designed fusion and selection channel is used to fuse the multi-modality feature and make the decision of the feature selection. Then, the proposed mechanism of coordinate sharing with padding integrates the multi-task of segmentation and detection so that it enables multi-task to perform united adversarial learning in one discriminator. Lastly, an innovative multi-phase radiomics guided discriminator exploits the clear and specific tumor information to improve the multi-task performance via the adversarial learning strategy. The UAL is validated in corresponding multi-modality NCMRI (i.e. T1FS pre-contrast MRI, T2FS MRI, and DWI) and three phases contrast-enhanced MRI of 255 clinical subjects. The experiments show that UAL has great potential in the clinical diagnosis of liver tumors.

Temporal abstraction in reinforcement learning is the ability of an agent to learn and use high-level behaviors, called options. The option-critic architecture provides a gradient-based end-to-end learning method to construct options. We propose an attention-based extension to this framework, which enables the agent to learn to focus different options on different aspects of the observation space. We show that this leads to behaviorally diverse options which are also capable of state abstraction, and prevents the degeneracy problems of option domination and frequent option switching that occur in option-critic, while achieving a similar sample complexity. We also demonstrate the more efficient, interpretable, and reusable nature of the learned options in comparison with option-critic, through different transfer learning tasks. Experimental results in a relatively simple four-rooms environment and the more complex ALE (Arcade Learning Environment) showcase the efficacy of our approach.