Curriculum learning has been a quiet yet crucial component of many of the high-profile successes of reinforcement learning. Despite this, none of the major reinforcement learning libraries directly support curriculum learning or include curriculum learning implementations. These methods can improve the capabilities and robustness of RL agents, but often require significant, complex changes to agent training code. We introduce Syllabus, a library for training RL agents with curriculum learning, as a solution to this problem. Syllabus provides a universal API for curriculum learning algorithms, implementations of popular curriculum learning methods, and infrastructure for easily integrating them with distributed training code written in nearly any RL library. Syllabus provides a minimal API for each of the core components of curriculum learning, dramatically simplifying the process of designing new algorithms and applying existing algorithms to new environments. We demonstrate that the same Syllabus code can be used to train agents written in multiple different RL libraries on numerous domains. In doing so, we present the first examples of curriculum learning in NetHack and Neural MMO, two of the premier challenges for single-agent and multi-agent RL respectively, achieving strong results compared to state of the art baselines.
In decentralized multi-agent reinforcement learning, agents learning in isolation can lead to relative over-generalization (RO), where optimal joint actions are undervalued in favor of suboptimal ones. This hinders effective coordination in cooperative tasks, as agents tend to choose actions that are individually rational but collectively suboptimal. To address this issue, we introduce MaxMax Q-Learning (MMQ), which employs an iterative process of sampling and evaluating potential next states, selecting those with maximal Q-values for learning. This approach refines approximations of ideal state transitions, aligning more closely with the optimal joint policy of collaborating agents. We provide theoretical analysis supporting MMQ's potential and present empirical evaluations across various environments susceptible to RO. Our results demonstrate that MMQ frequently outperforms existing baselines, exhibiting enhanced convergence and sample efficiency.
Differential privacy (DP) has recently been introduced into episodic reinforcement learning (RL) to formally address user privacy concerns in personalized services. Previous work mainly focuses on two trust models of DP: the central model, where a central agent is responsible for protecting users' sensitive data, and the (stronger) local model, where the protection occurs directly on the user side. However, they either require a trusted central agent or incur a significantly higher privacy cost, making it unsuitable for many scenarios. This work introduces a trust model stronger than the central model but with a lower privacy cost than the local model, leveraging the emerging \emph{shuffle} model of privacy. We present the first generic algorithm for episodic RL under the shuffle model, where a trusted shuffler randomly permutes a batch of users' data before sending it to the central agent. We then instantiate the algorithm using our proposed shuffle Privatizer, relying on a shuffle private binary summation mechanism. Our analysis shows that the algorithm achieves a near-optimal regret bound comparable to that of the centralized model and significantly outperforms the local model in terms of privacy cost.
We search for efficient disentanglers on random Clifford circuits of two-qubit gates arranged in a brick-wall pattern, using the proximal policy optimization (PPO) algorithm \cite{schulman2017proximalpolicyoptimizationalgorithms}. Disentanglers are defined as a set of projective measurements inserted between consecutive entangling layers. An efficient disentangler is a set of projective measurements that minimize the averaged von Neumann entropy of the final state with the least number of total projections possible. The problem is naturally amenable to reinforcement learning techniques by taking the binary matrix representing the projective measurements along the circuit as our state, and actions as bit flipping operations on this binary matrix that add or delete measurements at specified locations. We give rewards to our agent dependent on the averaged von Neumann entropy of the final state and the configuration of measurements, such that the agent learns the optimal policy that will take him from the initial state of no measurements to the optimal measurement state that minimizes the entanglement entropy. Our results indicate that the number of measurements required to disentangle a random quantum circuit is drastically less than the numerical results of measurement-induced phase transition papers. Additionally, the reinforcement learning procedure enables us to characterize the pattern of optimal disentanglers, which is not possible in the works of measurement-induced phase transitions.
Batch reinforcement learning enables policy learning without direct interaction with the environment during training, relying exclusively on previously collected sets of interactions. This approach is, therefore, well-suited for high-risk and cost-intensive applications, such as industrial control. Learned policies are commonly restricted to act in a similar fashion as observed in the batch. In a real-world scenario, learned policies are deployed in the industrial system, inevitably leading to the collection of new data that can subsequently be added to the existing recording. The process of learning and deployment can thus take place multiple times throughout the lifespan of a system. In this work, we propose to exploit this iterative nature of applying offline reinforcement learning to guide learned policies towards efficient and informative data collection during deployment, leading to continuous improvement of learned policies while remaining within the support of collected data. We present an algorithmic methodology for iterative batch reinforcement learning based on ensemble-based model-based policy search, augmented with safety and, importantly, a diversity criterion.
We propose a new risk sensitive reinforcement learning approach for the dynamic hedging of options. The approach focuses on the minimization of the tail risk of the final P&L of the seller of an option. Different from most existing reinforcement learning approaches that require a parametric model of the underlying asset, our approach can learn the optimal hedging strategy directly from the historical market data without specifying a parametric model; in addition, the learned optimal hedging strategy is contract-unified, i.e., it applies to different options contracts with different initial underlying prices, strike prices, and maturities. Our approach extends existing reinforcement learning methods by learning the tail risk measures of the final hedging P&L and the optimal hedging strategy at the same time. We carry out comprehensive empirical study to show that, in the out-of-sample tests, the proposed reinforcement learning hedging strategy can obtain statistically significantly lower tail risk and higher mean of the final P&L than delta hedging methods.
Efficient multiple setpoint tracking can enable advanced biotechnological applications, such as maintaining desired population levels in co-cultures for optimal metabolic division of labor. In this study, we employ reinforcement learning as a control method for population setpoint tracking in co-cultures, focusing on policy-gradient techniques where the control policy is parameterized by neural networks. However, achieving accurate tracking across multiple setpoints is a significant challenge in reinforcement learning, as the agent must effectively balance the contributions of various setpoints to maximize the expected system performance. Traditional return functions, such as those based on a quadratic cost, often yield suboptimal performance due to their inability to efficiently guide the agent toward the simultaneous satisfaction of all setpoints. To overcome this, we propose a novel return function that rewards the simultaneous satisfaction of multiple setpoints and diminishes overall reward gains otherwise, accounting for both stage and terminal system performance. This return function includes parameters to fine-tune the desired smoothness and steepness of the learning process. We demonstrate our approach considering an $\textit{Escherichia coli}$ co-culture in a chemostat with optogenetic control over amino acid synthesis pathways, leveraging auxotrophies to modulate growth.
Recent advances in deep learning suggest that we need to maximize and minimize two different kinds of information simultaneously. The Information Max-Min (IMM) method has been used in deep learning, reinforcement learning, and maximum entropy control. Shannon's information rate-distortion function is the theoretical basis of Minimizing Mutual Information (MMI) and data compression, but it is not enough to solve the IMM problem. The author has proposed the semantic information G theory (i.e., Shannon-Lu theory), including the semantic information G measure and the information rate fidelity function R(G) (R is the MMI for the given G of semantic mutual information). The parameter solution of the R(G) function provides a general method to improve the information efficiency, G/R. This paper briefly introduces the semantic information G measure and the parametric solution of the R(G) function. Two examples reveal that the parametric solution can help us optimize range control with the tradeoff between purposiveness (i.e., semantic mutual information) and information efficiency. It seems that the R(G) function can serve as the theoretical basis of IMM methods, but we still need further research in combination with deep learning, reinforcement learning, and constraint control.
Innate values describe agents' intrinsic motivations, which reflect their inherent interests and preferences for pursuing goals and drive them to develop diverse skills that satisfy their various needs. Traditional reinforcement learning (RL) is learning from interaction based on the feedback rewards of the environment. However, in real scenarios, the rewards are generated by agents' innate value systems, which differ vastly from individuals based on their needs and requirements. In other words, considering the AI agent as a self-organizing system, developing its awareness through balancing internal and external utilities based on its needs in different tasks is a crucial problem for individuals learning to support others and integrate community with safety and harmony in the long term. To address this gap, we propose a new RL model termed innate-values-driven RL (IVRL) based on combined motivations' models and expected utility theory to mimic its complex behaviors in the evolution through decision-making and learning. Then, we introduce two IVRL-based models: IV-DQN and IV-A2C. By comparing them with benchmark algorithms such as DQN, DDQN, A2C, and PPO in the Role-Playing Game (RPG) reinforcement learning test platform VIZDoom, we demonstrated that the IVRL-based models can help the agent rationally organize various needs, achieve better performance effectively.
Code-mixing(CM) or code-switching(CSW) refers to the juxtaposition of linguistic units from two or more languages during the conversation or sometimes even a single utterance. Code-mixing introduces unique challenges in daily life, such as syntactic mismatches and semantic blending, that are rarely encountered in monolingual settings. Large language models (LLMs) have revolutionized the field of natural language processing (NLP) by offering unprecedented capabilities in understanding human languages. However, the effectiveness of current state-of-the-art multilingual LLMs has not yet been fully explored in the CM scenario. To fill this gap, we first benchmark the performance of multilingual LLMs on various code-mixing NLP tasks. Then we propose to improve the multilingual LLMs' ability to understand code-mixing through reinforcement learning from human feedback (RLHF) and code-mixed machine translation tasks. Given the high-cost and time-consuming preference labeling procedure, we improve this by utilizing LLMs as annotators to perform the reinforcement learning from AI feedback (RLAIF). The experiments show the effectiveness of the proposed method.
This paper proposes a novel reinforcement learning framework to address the Liner Shipping Network Design Problem (LSNDP), a challenging combinatorial optimization problem focused on designing cost-efficient maritime shipping routes. Traditional methods for solving the LSNDP typically involve decomposing the problem into sub-problems, such as network design and multi-commodity flow, which are then tackled using approximate heuristics or large neighborhood search (LNS) techniques. In contrast, our approach employs a model-free reinforcement learning algorithm on the network design, integrated with a heuristic-based multi-commodity flow solver, to produce competitive results on the publicly available LINERLIB benchmark. Additionally, our method also demonstrates generalization capabilities by producing competitive solutions on the benchmark instances after training on perturbed instances.
The indoor environment significantly impacts human health and well-being; enhancing health and reducing energy consumption in these settings is a central research focus. With the advancement of Information and Communication Technology (ICT), recommendation systems and reinforcement learning (RL) have emerged as promising approaches to induce behavioral changes to improve the indoor environment and energy efficiency of buildings. This study aims to employ text mining and Natural Language Processing (NLP) techniques to thoroughly examine the connections among these approaches in the context of human-building interaction and occupant context-aware support. The study analyzed 27,595 articles from the ScienceDirect database, revealing extensive use of recommendation systems and RL for space optimization, location recommendations, and personalized control suggestions. Furthermore, this review underscores the vast potential for expanding recommender systems and RL applications in buildings and indoor environments. Fields ripe for innovation include predictive maintenance, building-related product recommendation, and optimization of environments tailored for specific needs, such as sleep and productivity enhancements based on user feedback. The study also notes the limitations of the method in capturing subtle academic nuances. Future improvements could involve integrating and fine-tuning pre-trained language models to better interpret complex texts.
Autonomous robots collaboratively exploring an unknown environment is still an open problem. The problem has its roots in coordination among non-stationary agents, each with only a partial view of information. The problem is compounded when the multiple robots must completely explore the environment. In this paper, we introduce Backtrack Assisted Multi-Agent Exploration using Reinforcement Learning (BAMAX), a method for collaborative exploration in multi-agent systems which attempts to explore an entire virtual environment. As in the name, BAMAX leverages backtrack assistance to enhance the performance of agents in exploration tasks. To evaluate BAMAX against traditional approaches, we present the results of experiments conducted across multiple hexagonal shaped grids sizes, ranging from 10x10 to 60x60. The results demonstrate that BAMAX outperforms other methods in terms of faster coverage and less backtracking across these environments.
Reinforcement Learning (RL) is a rapidly growing area of machine learning that finds its application in a broad range of domains, from finance and healthcare to robotics and gaming. Compared to other machine learning techniques, RL agents learn from their own experiences using trial and error, and improve their performance over time. However, assessing RL models can be challenging, which makes it difficult to interpret their behaviour. While reward is a widely used metric to evaluate RL models, it may not always provide an accurate measure of training performance. In some cases, the reward may seem increasing while the model's performance is actually decreasing, leading to misleading conclusions about the effectiveness of the training. To overcome this limitation, we have developed RLInspect - an interactive visual analytic tool, that takes into account different components of the RL model - state, action, agent architecture and reward, and provides a more comprehensive view of the RL training. By using RLInspect, users can gain insights into the model's behaviour, identify issues during training, and potentially correct them effectively, leading to a more robust and reliable RL system.
Reinforcement learning from human feedback (RLHF) provides a paradigm for aligning large language models (LLMs) with human preferences. This involves the initial training of a reward model based on pairwise human feedback. The reward model is subsequently utilized in reinforcement learning to assess the scores of each generated sentence as a whole, further guiding the optimization of LLMs. However, current approaches have a significant shortcoming: \emph{They allocate a single, sparse, and delayed reward to an entire sequence of output}. This may overlook some significant individual contributions of each token towards the desired outcome. To overcome this limitation, our paper proposes a novel reward redistribution method called R3HF, which facilitates a more fine-grained, token-level reward allocation. Specifically, our method treats the reward prediction task of the reward model as a regression problem. As a result, the redistributed rewards are computed by evaluating the specific contribution of each token to the reward model's output. This detailed approach improves the model's understanding of language nuances, leading to more precise enhancements in its performance. Our method is crafted to integrate seamlessly with most current techniques while incurring minimal computational costs. Through comprehensive experiments across diverse datasets and tasks, we have verified the effectiveness and superiority of our approach.
Non-prehensile manipulation, such as pushing objects to a desired target position, is an important skill for robots to assist humans in everyday situations. However, the task is challenging due to the large variety of objects with different and sometimes unknown physical properties, such as shape, size, mass, and friction. This can lead to the object overshooting its target position, requiring fast corrective movements of the robot around the object, especially in cases where objects need to be precisely pushed. In this paper, we improve the state-of-the-art by introducing a new memory-based vision-proprioception RL model to push objects more precisely to target positions using fewer corrective movements.
Coordination in traffic signal control is crucial for managing congestion in urban networks. Existing pressure-based control methods focus only on immediate upstream links, leading to suboptimal green time allocation and increased network delays. However, effective signal control inherently requires coordination across a broader spatial scope, as the effect of upstream traffic should influence signal control decisions at downstream intersections, impacting a large area in the traffic network. Although agent communication using neural network-based feature extraction can implicitly enhance spatial awareness, it significantly increases the learning complexity, adding an additional layer of difficulty to the challenging task of control in deep reinforcement learning. To address the issue of learning complexity and myopic traffic pressure definition, our work introduces a novel concept based on Markov chain theory, namely \textit{multi-hop upstream pressure}, which generalizes the conventional pressure to account for traffic conditions beyond the immediate upstream links. This farsighted and compact metric informs the deep reinforcement learning agent to preemptively clear the multi-hop upstream queues, guiding the agent to optimize signal timings with a broader spatial awareness. Simulations on synthetic and realistic (Toronto) scenarios demonstrate controllers utilizing multi-hop upstream pressure significantly reduce overall network delay by prioritizing traffic movements based on a broader understanding of upstream congestion.
Distributionally robust offline reinforcement learning (RL) aims to find a policy that performs the best under the worst environment within an uncertainty set using an offline dataset collected from a nominal model. While recent advances in robust RL focus on Markov decision processes (MDPs), robust non-Markovian RL is limited to planning problem where the transitions in the uncertainty set are known. In this paper, we study the learning problem of robust offline non-Markovian RL. Specifically, when the nominal model admits a low-rank structure, we propose a new algorithm, featuring a novel dataset distillation and a lower confidence bound (LCB) design for robust values under different types of the uncertainty set. We also derive new dual forms for these robust values in non-Markovian RL, making our algorithm more amenable to practical implementation. By further introducing a novel type-I concentrability coefficient tailored for offline low-rank non-Markovian decision processes, we prove that our algorithm can find an $\epsilon$-optimal robust policy using $O(1/\epsilon^2)$ offline samples. Moreover, we extend our algorithm to the case when the nominal model does not have specific structure. With a new type-II concentrability coefficient, the extended algorithm also enjoys polynomial sample efficiency under all different types of the uncertainty set.
Sepsis is a life-threatening condition defined by end-organ dysfunction due to a dysregulated host response to infection. Although the Surviving Sepsis Campaign has launched and has been releasing sepsis treatment guidelines to unify and normalize the care for sepsis patients, it has been reported in numerous studies that disparities in care exist across the trajectory of patient stay in the emergency department and intensive care unit. Here, we apply a number of reinforcement learning techniques including behavioral cloning, imitation learning, and inverse reinforcement learning, to learn the optimal policy in the management of septic patient subgroups using expert demonstrations. Then we estimate the counterfactual optimal policies by applying the model to another subset of unseen medical populations and identify the difference in cure by comparing it to the real policy. Our data comes from the sepsis cohort of MIMIC-IV and the clinical data warehouses of the Mass General Brigham healthcare system. The ultimate objective of this work is to use the optimal learned policy function to estimate the counterfactual treatment policy and identify deviations across sub-populations of interest. We hope this approach would help us identify any disparities in care and also changes in cure in response to the publication of national sepsis treatment guidelines.
Offline Reinforcement Learning (RL) suffers from the extrapolation error and value overestimation. From a generalization perspective, this issue can be attributed to the over-generalization of value functions or policies towards out-of-distribution (OOD) actions. Significant efforts have been devoted to mitigating such generalization, and recent in-sample learning approaches have further succeeded in entirely eschewing it. Nevertheless, we show that mild generalization beyond the dataset can be trusted and leveraged to improve performance under certain conditions. To appropriately exploit generalization in offline RL, we propose Doubly Mild Generalization (DMG), comprising (i) mild action generalization and (ii) mild generalization propagation. The former refers to selecting actions in a close neighborhood of the dataset to maximize the Q values. Even so, the potential erroneous generalization can still be propagated, accumulated, and exacerbated by bootstrapping. In light of this, the latter concept is introduced to mitigate the generalization propagation without impeding the propagation of RL learning signals. Theoretically, DMG guarantees better performance than the in-sample optimal policy in the oracle generalization scenario. Even under worst-case generalization, DMG can still control value overestimation at a certain level and lower bound the performance. Empirically, DMG achieves state-of-the-art performance across Gym-MuJoCo locomotion tasks and challenging AntMaze tasks. Moreover, benefiting from its flexibility in both generalization aspects, DMG enjoys a seamless transition from offline to online learning and attains strong online fine-tuning performance.