In this article, we aim to provide a literature review of different formulations and approaches to continual reinforcement learning (RL), also known as lifelong or non-stationary RL. We begin by discussing our perspective on why RL is a natural fit for studying continual learning. We then provide a taxonomy of different continual RL formulations by mathematically characterizing two key properties of non-stationarity, namely, the scope and driver non-stationarity. This offers a unified view of various formulations. Next, we review and present a taxonomy of continual RL approaches. We go on to discuss evaluation of continual RL agents, providing an overview of benchmarks used in the literature and important metrics for understanding agent performance. Finally, we highlight open problems and challenges in bridging the gap between the current state of continual RL and findings in neuroscience. While still in its early days, the study of continual RL has the promise to develop better incremental reinforcement learners that can function in increasingly realistic applications where non-stationarity plays a vital role. These include applications such as those in the fields of healthcare, education, logistics, and robotics.

You have an environment, a PyTorch model, and an RL framework that are designed to work together but don't. PufferLib is a wrapper layer that provide better compatibility between Gym / PettingZoo environments and standard reinforcement learning frameworks. You write a native PyTorch network and a short binding for your environment; PufferLib takes care of the rest. We plan to add additional bindings in the future. These mainly provide a wrapper utility that creates a framework-compliant network from a raw PyTorch model.

In the AI realm, reinforcement learning (RL) is lauded for good reasons. It is one of the most important advancements towards enabling general AI. But outside of popular interest, some researchers question whether it is the correct way to train machines in order to move forward. The technique has often been described as "the first computational theory of intelligence" by scientists. One of the players that have made it to the top of the reinforcement learning leaderboard is DeepMind, a London-based research firm.

Microswimmers can acquire information on the surrounding fluid by sensing mechanical queues. They can then navigate in response to these signals. We analyse this navigation by combining deep reinforcement learning with direct numerical simulations to resolve the hydrodynamics. We study how local and non-local information can be used to train a swimmer to achieve particular swimming tasks in a non-uniform flow field, in particular a zig-zag shear flow. The swimming tasks are (1) learning how to swim in the vorticity direction, (2) the shear-gradient direction, and (3) the shear flow direction. We find that access to lab frame information on the swimmer's instantaneous orientation is all that is required in order to reach the optimal policy for (1,2). However, information on both the translational and rotational velocities seem to be required to achieve (3). Inspired by biological microorganisms we also consider the case where the swimmers sense local information, i.e. surface hydrodynamic forces, together with a signal direction. This might correspond to gravity or, for micro-organisms with light sensors, a light source. In this case, we show that the swimmer can reach a comparable level of performance as a swimmer with access to lab frame variables. We also analyse the role of different swimming modes, i.e. pusher, puller, and neutral swimmers.

Visual reinforcement learning (RL), which makes decisions directly from high-dimensional visual inputs, has demonstrated significant potential in various domains. However, deploying visual RL techniques in the real world remains challenging due to their low sample efficiency and large generalization gaps. To tackle these obstacles, data augmentation (DA) has become a widely used technique in visual RL for acquiring sample-efficient and generalizable policies by diversifying the training data. This survey aims to provide a timely and essential review of DA techniques in visual RL in recognition of the thriving development in this field. In particular, we propose a unified framework for analyzing visual RL and understanding the role of DA in it. We then present a principled taxonomy of the existing augmentation techniques used in visual RL and conduct an in-depth discussion on how to better leverage augmented data in different scenarios. Moreover, we report a systematic empirical evaluation of DA-based techniques in visual RL and conclude by highlighting the directions for future research. As the first comprehensive survey of DA in visual RL, this work is expected to offer valuable guidance to this emerging field.

Although Reinforcement Learning (RL) has shown impressive results in games and simulation, real-world application of RL suffers from its instability under changing environment conditions and hyperparameters. We give a first impression of the extent of this instability by showing that the hyperparameters found by automatic hyperparameter optimization (HPO) methods are not only dependent on the problem at hand, but even on how well the state describes the environment dynamics. Specifically, we show that agents in contextual RL require different hyperparameters if they are shown how environmental factors change. In addition, finding adequate hyperparameter configurations is not equally easy for both settings, further highlighting the need for research into how hyperparameters influence learning and generalization in RL.

In the context of an efficient network traffic engineering process where the network continuously measures a new traffic matrix and updates the set of paths in the network, an automated process is required to quickly and efficiently identify when and what set of paths should be used. Unfortunately, the burden of finding the optimal solution for the network updating process in each given time interval is high since the computation complexity of optimization approaches using linear programming increases significantly as the size of the network increases. In this paper, we use deep reinforcement learning to derive a data-driven algorithm that does the path selection in the network considering the overhead of route computation and path updates. Our proposed scheme leverages information about past network behavior to identify a set of robust paths to be used for multiple future time intervals to avoid the overhead of updating the forwarding behavior of routers frequently. We compare the results of our approach to other traffic engineering solutions through extensive simulations across real network topologies. Our results demonstrate that our scheme fares well by a factor of 40% with respect to reducing link utilization compared to traditional TE schemes such as ECMP. Our scheme provides a slightly higher link utilization (around 25%) compared to schemes that only minimize link utilization and do not care about path updating overhead.

Many machine-learning algorithms work with the assumption that all features have been observed and available during training and testing times or the missing data are disregarded as unacquired. Feature acquisition, a process in which further relevant data are acquired at variable costs, addresses this assumption to more closely align with some real-world applications, Huang [3]. For medical diagnostic tasks, from the basis of incomplete features, doctors sequentially obtain additional test results until they obtain sufficient information to make adequate diagnoses of the patients. Determining which features to acquire is dependent on the previous diagnostic observations and the sequence at which the features are obtained can vary from patient to patient. Although accurate diagnoses are more likely with additional features, acquiring them incurs variable costs and is balanced with the improvement in performance, Melville [1]. Previous studies on the feature acquisition problem address the trade-off between acquisition costs and performance improvement and the sequential decision making process, and are categorized into non-reinforcement learning and reinforcement learning (RL) approaches. Non-RL approaches focus on selecting the most informative features to acquire based on their utility values. These methods, Melville [1], desJardins [2], and Huang [3], estimate the expected utility of a feature for improving the model performance and acquire the feature with maximum expected utility.

Steering language generation towards objectives or away from undesired content has been a long-standing goal in utilizing language models (LM). Recent work has demonstrated reinforcement learning and weighted decoding as effective approaches to achieve a higher level of language control and quality with pros and cons. In this work, we propose a novel critic decoding method for controlled language generation (CriticControl) that combines the strengths of reinforcement learning and weighted decoding. Specifically, we adopt the actor-critic framework to train an LM-steering critic from non-differentiable reward models. And similar to weighted decoding, our method freezes the language model and manipulates the output token distribution using called critic, improving training efficiency and stability. Evaluation of our method on three controlled generation tasks, namely topic control, sentiment control, and detoxification, shows that our approach generates more coherent and well-controlled texts than previous methods. In addition, CriticControl demonstrates superior generalization ability in zero-shot settings. Human evaluation studies also corroborate our findings.

Sentence summarization shortens given texts while maintaining core contents of the texts. Unsupervised approaches have been studied to summarize texts without human-written summaries. However, recent unsupervised models are extractive, which remove words from texts and thus they are less flexible than abstractive summarization. In this work, we devise an abstractive model based on reinforcement learning without ground-truth summaries. We formulate the unsupervised summarization based on the Markov decision process with rewards representing the summary quality. To further enhance the summary quality, we develop a multi-summary learning mechanism that generates multiple summaries with varying lengths for a given text, while making the summaries mutually enhance each other. Experimental results show that the proposed model substantially outperforms both abstractive and extractive models, yet frequently generating new words not contained in input texts.