We consider reinforcement learning (RL) for a class of problems with bagged decision times. A bag contains a finite sequence of consecutive decision times. The transition dynamics are non-Markovian and non-stationary within a bag. All actions within a bag jointly impact a single reward, observed at the end of the bag. For example, in mobile health, multiple activity suggestions in a day collectively affect a user's daily commitment to being active. Our goal is to develop an online RL algorithm to maximize the discounted sum of the bag-specific rewards. To handle non-Markovian transitions within a bag, we utilize an expert-provided causal directed acyclic graph (DAG). Based on the DAG, we construct states as a dynamical Bayesian sufficient statistic of the observed history, which results in Markov state transitions within and across bags. We then formulate this problem as a periodic Markov decision process (MDP) that allows non-stationarity within a period. An online RL algorithm based on Bellman equations for stationary MDPs is generalized to handle periodic MDPs. We show that our constructed state achieves the maximal optimal value function among all state constructions for a periodic MDP. Finally, we evaluate the proposed method on testbed variants built from real data in a mobile health clinical trial.
Traditional methods for computing equilibria in auctions become computationally intractable as auction complexity increases, particularly in multi-item and dynamic auctions. This paper introduces a self-play based reinforcement learning approach that employs advanced algorithms such as Proximal Policy Optimization and Neural Fictitious Self-Play to approximate Bayes-Nash equilibria. This framework allows for continuous action spaces, high-dimensional information states, and delayed payoffs. Through self-play, these algorithms can learn robust and near-optimal bidding strategies in auctions with known equilibria, including those with symmetric and asymmetric valuations, private and interdependent values, and multi-round auctions.
This paper considers an online reinforcement learning algorithm that leverages pre-collected data (passive memory) from the environment for online interaction. We show that using passive memory improves performance and further provide theoretical guarantees for regret that turns out to be near-minimax optimal. Results show that the quality of passive memory determines sub-optimality of the incurred regret. The proposed approach and results hold in both continuous and discrete state-action spaces.
Deep Reinforcement learning (DRL) is used to enable autonomous navigation in unknown environments. Most research assume perfect sensor data, but real-world environments may contain natural and artificial sensor noise and denial. Here, we present a benchmark of both well-used and emerging DRL algorithms in a navigation task with configurable sensor denial effects. In particular, we are interested in comparing how different DRL methods (e.g. model-free PPO vs. model-based DreamerV3) are affected by sensor denial. We show that DreamerV3 outperforms other methods in the visual end-to-end navigation task with a dynamic goal - and other methods are not able to learn this. Furthermore, DreamerV3 generally outperforms other methods in sensor-denied environments. In order to improve robustness, we use adversarial training and demonstrate an improved performance in denied environments, although this generally comes with a performance cost on the vanilla environments. We anticipate this benchmark of different DRL methods and the usage of adversarial training to be a starting point for the development of more elaborate navigation strategies that are capable of dealing with uncertain and denied sensor readings.
In this paper, we consider a transfer reinforcement learning problem involving agents with different action spaces. Specifically, for any new unseen task, the goal is to use a successful demonstration of this task by an expert agent in its action space to enable a learner agent learn an optimal policy in its own different action space with fewer samples than those required if the learner was learning on its own. Existing transfer learning methods across different action spaces either require handcrafted mappings between those action spaces provided by human experts, which can induce bias in the learning procedure, or require the expert agent to share its policy parameters with the learner agent, which does not generalize well to unseen tasks. In this work, we propose a method that learns a subgoal mapping between the expert agent policy and the learner agent policy. Since the expert agent and the learner agent have different action spaces, their optimal policies can have different subgoal trajectories. We learn this subgoal mapping by training a Long Short Term Memory (LSTM) network for a distribution of tasks and then use this mapping to predict the learner subgoal sequence for unseen tasks, thereby improving the speed of learning by biasing the agent's policy towards the predicted learner subgoal sequence. Through numerical experiments, we demonstrate that the proposed learning scheme can effectively find the subgoal mapping underlying the given distribution of tasks. Moreover, letting the learner agent imitate the expert agent's policy with the learnt subgoal mapping can significantly improve the sample efficiency and training time of the learner agent in unseen new tasks.
Deep reinforcement learning (RL) agents rely on shortcut learning, preventing them from generalizing to slightly different environments. To address this problem, symbolic method, that use object-centric states, have been developed. However, comparing these methods to deep agents is not fair, as these last operate from raw pixel-based states. In this work, we instantiate the symbolic SCoBots framework. SCoBots decompose RL tasks into intermediate, interpretable representations, culminating in action decisions based on a comprehensible set of object-centric relational concepts. This architecture aids in demystifying agent decisions. By explicitly learning to extract object-centric representations from raw states, object-centric RL, and policy distillation via rule extraction, this work places itself within the neurosymbolic AI paradigm, blending the strengths of neural networks with symbolic AI. We present the first implementation of an end-to-end trained SCoBot, separately evaluate of its components, on different Atari games. The results demonstrate the framework's potential to create interpretable and performing RL systems, and pave the way for future research directions in obtaining end-to-end interpretable RL agents.
With the rapid development of the internet, the richness of User-Generated Contentcontinues to increase, making Multimodal Aspect-Based Sentiment Analysis (MABSA) a research hotspot. Existing studies have achieved certain results in MABSA, but they have not effectively addressed the analytical challenges in scenarios where multiple entities and sentiments coexist. This paper innovatively introduces Large Language Models (LLMs) for event decomposition and proposes a reinforcement learning framework for Multimodal Aspect-based Sentiment Analysis (MABSA-RL) framework. This framework decomposes the original text into a set of events using LLMs, reducing the complexity of analysis, introducing reinforcement learning to optimize model parameters. Experimental results show that MABSA-RL outperforms existing advanced methods on two benchmark datasets. This paper provides a new research perspective and method for multimodal aspect-level sentiment analysis.
In this paper, we study an inverse reinforcement learning problem that involves learning the reward function of a learning agent using trajectory data collected while this agent is learning its optimal policy. To address this problem, we propose an inverse reinforcement learning method that allows us to estimate the policy parameters of the learning agent which can then be used to estimate its reward function. Our method relies on a new variant of the behavior cloning algorithm, which we call bundle behavior cloning, and uses a small number of trajectories generated by the learning agent's policy at different points in time to learn a set of policies that match the distribution of actions observed in the sampled trajectories. We then use the cloned policies to train a neural network model that estimates the reward function of the learning agent. We provide a theoretical analysis to show a complexity result on bound guarantees for our method that beats standard behavior cloning as well as numerical experiments for a reinforcement learning problem that validate the proposed method.
To coordinate the interests of operator and users in a microgrid under complex and changeable operating conditions, this paper proposes a microgrid scheduling model considering the thermal flexibility of thermostatically controlled loads and demand response by leveraging physical informed-inspired deep reinforcement learning (DRL) based bi-level programming. To overcome the non-convex limitations of karush-kuhn-tucker (KKT)-based methods, a novel optimization solution method based on DRL theory is proposed to handle the bi-level programming through alternate iterations between levels. Specifically, by combining a DRL algorithm named asynchronous advantage actor-critic (A3C) and automated machine learning-prioritized experience replay (AutoML-PER) strategy to improve the generalization performance of A3C to address the above problems, an improved A3C algorithm, called AutoML-PER-A3C, is designed to solve the upper-level problem; while the DOCPLEX optimizer is adopted to address the lower-level problem. In this solution process, AutoML is used to automatically optimize hyperparameters and PER improves learning efficiency and quality by extracting the most valuable samples. The test results demonstrate that the presented approach manages to reconcile the interests between multiple stakeholders in MG by fully exploiting various flexibility resources. Furthermore, in terms of economic viability and computational efficiency, the proposal vastly exceeds other advanced reinforcement learning methods.
Given its intuitive nature, many Cloud providers opt for threshold-based data replication to enable automatic resource scaling. However, setting thresholds effectively needs human intervention to calibrate thresholds for each metric and requires a deep knowledge of current workload trends, which can be challenging to achieve. Reinforcement learning is used in many areas related to the Cloud Computing, and it is a promising field to get automatic data replication strategies. In this work, we survey data replication strategies and data scaling based on reinforcement learning (RL).
Reinforcement Learning from Human Feedback (RLHF) has greatly improved the performance of modern Large Language Models (LLMs). The RLHF process is resource-intensive and technically challenging, generally requiring a large collection of human preference labels over model-generated outputs. Reinforcement Learning from AI Feedback (RLAIF) addresses this data collection challenge by leveraging synthetic preferences generated by an LLM. However, recent work has shown that synthetic preferences labels may not align well with human preference judgments. To address this, we propose a hybrid approach that unifies RLHF and RLAIF methodologies. We introduce GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments. Empirically, we show that zero-shot LLM-based judgments under-perform compared to Bradley-Terry reward models on in-distribution tasks (between 9-36%). In contrast, GenRM achieves in-distribution accuracy comparable to Bradley-Terry models, while significantly outperforming them on out-of-distribution tasks (between 10-45%). Moreover, GenRM surpasses the performance of using LLMs as judges on both in-distribution (by 9-31%) and out-of-distribution tasks (by 2- 6%). Our results show that combining the strengths of RLHF and RLAIF offers a promising approach for improving the quality of synthetic preference labels.
We study the problem of Non-Stationary Reinforcement Learning (NS-RL) without prior knowledge about the system's non-stationarity. A state-of-the-art, black-box algorithm, known as MASTER, is considered, with a focus on identifying the conditions under which it can achieve its stated goals. Specifically, we prove that MASTER's non-stationarity detection mechanism is not triggered for practical choices of horizon, leading to performance akin to a random restarting algorithm. Moreover, we show that the regret bound for MASTER, while being order optimal, stays above the worst-case linear regret until unreasonably large values of the horizon. To validate these observations, MASTER is tested for the special case of piecewise stationary multi-armed bandits, along with methods that employ random restarting, and others that use quickest change detection to restart. A simple, order optimal random restarting algorithm, that has prior knowledge of the non-stationarity is proposed as a baseline. The behavior of the MASTER algorithm is validated in simulations, and it is shown that methods employing quickest change detection are more robust and consistently outperform MASTER and other random restarting approaches.
This paper tackles the challenge of learning non-Markovian optimal execution strategies in dynamic financial markets. We introduce a novel actor-critic algorithm based on Deep Deterministic Policy Gradient (DDPG) to address this issue, with a focus on transient price impact modeled by a general decay kernel. Through numerical experiments with various decay kernels, we show that our algorithm successfully approximates the optimal execution strategy. Additionally, the proposed algorithm demonstrates adaptability to evolving market conditions, where parameters fluctuate over time. Our findings also show that modern reinforcement learning algorithms can provide a solution that reduces the need for frequent and inefficient human intervention in optimal execution tasks.
Reinforcement learning (RL) often necessitates a meticulous Markov Decision Process (MDP) design tailored to each task. This work aims to address this challenge by proposing a systematic approach to behavior synthesis and control for multi-contact loco-manipulation tasks, such as navigating spring-loaded doors and manipulating heavy dishwashers. We define a task-independent MDP to train RL policies using only a single demonstration per task generated from a model-based trajectory optimizer. Our approach incorporates an adaptive phase dynamics formulation to robustly track the demonstrations while accommodating dynamic uncertainties and external disturbances. We compare our method against prior motion imitation RL works and show that the learned policies achieve higher success rates across all considered tasks. These policies learn recovery maneuvers that are not present in the demonstration, such as re-grasping objects during execution or dealing with slippages. Finally, we successfully transfer the policies to a real robot, demonstrating the practical viability of our approach.
Large language models (LLMs) trained with Reinforcement Learning from Human Feedback (RLHF) have demonstrated remarkable capabilities, but their underlying reward functions and decision-making processes remain opaque. This paper introduces a novel approach to interpreting LLMs by applying inverse reinforcement learning (IRL) to recover their implicit reward functions. We conduct experiments on toxicity-aligned LLMs of varying sizes, extracting reward models that achieve up to 80.40% accuracy in predicting human preferences. Our analysis reveals key insights into the non-identifiability of reward functions, the relationship between model size and interpretability, and potential pitfalls in the RLHF process. We demonstrate that IRL-derived reward models can be used to fine-tune new LLMs, resulting in comparable or improved performance on toxicity benchmarks. This work provides a new lens for understanding and improving LLM alignment, with implications for the responsible development and deployment of these powerful systems.
In deep Reinforcement Learning (RL) models trained using gradient-based techniques, the choice of optimizer and its learning rate are crucial to achieving good performance: higher learning rates can prevent the model from learning effectively, while lower ones might slow convergence. Additionally, due to the non-stationarity of the objective function, the best-performing learning rate can change over the training steps. To adapt the learning rate, a standard technique consists of using decay schedulers. However, these schedulers assume that the model is progressively approaching convergence, which may not always be true, leading to delayed or premature adjustments. In this work, we propose dynamic Learning Rate for deep Reinforcement Learning (LRRL), a meta-learning approach that selects the learning rate based on the agent's performance during training. LRRL is based on a multi-armed bandit algorithm, where each arm represents a different learning rate, and the bandit feedback is provided by the cumulative returns of the RL policy to update the arms' probability distribution. Our empirical results demonstrate that LRRL can substantially improve the performance of deep RL algorithms for some tasks.
Human actions are based on the mental perception of the environment. Even when all the aspects of an environment are not visible, humans have an internal mental model that can generalize the partially visible scenes to fully constructed and connected views. This internal mental model uses learned abstract representations of spatial and temporal aspects of the environments encountered in the past. Artificial agents in reinforcement learning environments also benefit by learning a representation of the environment from experience. It provides the agent with viewpoints that are not directly visible to it, helping it make better policy decisions. It can also be used to predict the future state of the environment. This project explores learning the top-down view of an RL environment based on the artificial agent's first-person view observations with a generative adversarial network(GAN). The top-down view is useful as it provides a complete overview of the environment by building a map of the entire environment. It provides information about the objects' dimensions and shapes along with their relative positions with one another. Initially, when only a partial observation of the environment is visible to the agent, only a partial top-down view is generated. As the agent explores the environment through a set of actions, the generated top-down view becomes complete. This generated top-down view can assist the agent in deducing better policy decisions. The focus of the project is to learn the top-down view of an RL environment. It doesn't deal with any Reinforcement Learning task.
As the number of devices getting connected to the vehicular network grows exponentially, addressing the numerous challenges of effectively allocating spectrum in dynamic vehicular environment becomes increasingly difficult. Traditional methods may not suffice to tackle this issue. In vehicular networks safety critical messages are involved and it is important to implement an efficient spectrum allocation paradigm for hassle free communication as well as manage the congestion in the network. To tackle this, a Deep Q Network (DQN) model is proposed as a solution, leveraging its ability to learn optimal strategies over time and make decisions. The paper presents a few results and analyses, demonstrating the efficacy of the DQN model in enhancing spectrum sharing efficiency. Deep Reinforcement Learning methods for sharing spectrum in vehicular networks have shown promising outcomes, demonstrating the system's ability to adjust to dynamic communication environments. Both SARL and MARL models have exhibited successful rates of V2V communication, with the cumulative reward of the RL model reaching its maximum as training progresses.
We demonstrate that recent advances in reinforcement learning (RL) combined with simple architectural changes significantly improves generalization on the ProcGen benchmark. These changes are frame stacking, replacing 2D convolutional layers with 3D convolutional layers, and scaling up the number of convolutional kernels per layer. Experimental results using a single set of hyperparameters across all environments show a 37.9\% reduction in the optimality gap compared to the baseline (from 0.58 to 0.36). This performance matches or exceeds current state-of-the-art methods. The proposed changes are largely orthogonal and therefore complementary to the existing approaches for improving generalization in RL, and our results suggest that further exploration in this direction could yield substantial improvements in addressing generalization challenges in deep reinforcement learning.
Traditionally, model-based reinforcement learning (MBRL) methods exploit neural networks as flexible function approximators to represent $\textit{a priori}$ unknown environment dynamics. However, training data are typically scarce in practice, and these black-box models often fail to generalize. Modeling architectures that leverage known physics can substantially reduce the complexity of system-identification, but break down in the face of complex phenomena such as contact. We introduce a novel framework for learning semi-structured dynamics models for contact-rich systems which seamlessly integrates structured first principles modeling techniques with black-box auto-regressive models. Specifically, we develop an ensemble of probabilistic models to estimate external forces, conditioned on historical observations and actions, and integrate these predictions using known Lagrangian dynamics. With this semi-structured approach, we can make accurate long-horizon predictions with substantially less data than prior methods. We leverage this capability and propose Semi-Structured Reinforcement Learning ($\texttt{SSRL}$) a simple model-based learning framework which pushes the sample complexity boundary for real-world learning. We validate our approach on a real-world Unitree Go1 quadruped robot, learning dynamic gaits -- from scratch -- on both hard and soft surfaces with just a few minutes of real-world data. Video and code are available at: https://sites.google.com/utexas.edu/ssrl