Reinforcement Learning

In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators and achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads to predict the preference, one trained to correlate with length and the other trained to decorrelate with length and therefore focusing more on the actual content. We then discard the length head in RL to ignore the spurious length reward. Experiments demonstrate that our approach eliminates the reward correlation with length, and improves the obtained policy by a significant margin.

This paper contributes a new approach for distributional reinforcement learning which elucidates a clean separation of transition structure and reward in the learning process. Analogous to how the successor representation (SR) describes the expected consequences of behaving according to a given policy, our distributional successor measure (SM) describes the distributional consequences of this behaviour. We formulate the distributional SM as a distribution over distributions and provide theory connecting it with distributional and model-based reinforcement learning. Moreover, we propose an algorithm that learns the distributional SM from data by minimizing a two-level maximum mean discrepancy. Key to our method are a number of algorithmic techniques that are independently valuable for learning generative models of state. As an illustration of the usefulness of the distributional SM, we show that it enables zero-shot risk-sensitive policy evaluation in a way that was not previously possible.

In multi-agent reinforcement learning (MARL), effective exploration is critical, especially in sparse reward environments. Although introducing global intrinsic rewards can foster exploration in such settings, it often complicates credit assignment among agents. To address this difficulty, we propose Individual Contributions as intrinsic Exploration Scaffolds (ICES), a novel approach to motivate exploration by assessing each agent's contribution from a global view. In particular, ICES constructs exploration scaffolds with Bayesian surprise, leveraging global transition information during centralized training. These scaffolds, used only in training, help to guide individual agents towards actions that significantly impact the global latent state transitions. Additionally, ICES separates exploration policies from exploitation policies, enabling the former to utilize privileged global information during training. Extensive experiments on cooperative benchmark tasks with sparse rewards, including Google Research Football (GRF) and StarCraft Multi-agent Challenge (SMAC), demonstrate that ICES exhibits superior exploration capabilities compared with baselines. The code is publicly available at https://github.com/LXXXXR/ICES.

We study the problem of obtaining optimal and realistic prescriptions when using ReLU networks for data-driven decision-making. In this setting, the network is used to predict a quantity of interest and then is optimized to retrieve the decisions that maximize the quantity (e.g. find the best prices that maximize revenue). However, optimizing over-parameterized models often produces unrealistic prescriptions, far from the data manifold. This phenomenon is known as the Optimizer's Curse. To tackle this problem, we model the requirement for the resulting decisions to align with the data manifold as a tractable optimization constraint. This is achieved by reformulating the highly nonlinear Local Outlier Factor (LOF) metric as a single linear or quadratic constraint. To solve the problem efficiently for large networks, we propose an adaptive sampling algorithm that reduces the initial hard-to-solve optimization problem into a small number of significantly easier-to-solve problems by restricting the decision space to realistic polytopes, i.e. polytopes of the decision space that contain at least one realistic data point. Experiments on publicly available networks demonstrate the efficacy and scalability of our approach.

Considering generating samples with high rewards, we focus on optimizing deep neural networks parameterized stochastic differential equations (SDEs), the advanced generative models with high expressiveness, with policy gradient, the leading algorithm in reinforcement learning. Nevertheless, when applying policy gradients to SDEs, since the policy gradient is estimated on a finite set of trajectories, it can be ill-defined, and the policy behavior in data-scarce regions may be uncontrolled. This challenge compromises the stability of policy gradients and negatively impacts sample complexity. To address these issues, we propose constraining the SDE to be consistent with its associated perturbation process. Since the perturbation process covers the entire space and is easy to sample, we can mitigate the aforementioned problems. Our framework offers a general approach allowing for a versatile selection of policy gradient methods to effectively and efficiently train SDEs. We evaluate our algorithm on the task of structure-based drug design and optimize the binding affinity of generated ligand molecules. Our method achieves the best Vina score (-9.07) on the CrossDocked2020 dataset.

Second-order information is valuable for many applications but challenging to compute. Several works focus on computing or approximating Hessian diagonals, but even this simplification introduces significant additional costs compared to computing a gradient. In the absence of efficient exact computation schemes for Hessian diagonals, we revisit an early approximation scheme proposed by Becker and LeCun (1989, BL89), which has a cost similar to gradients and appears to have been overlooked by the community. We introduce HesScale, an improvement over BL89, which adds negligible extra computation. On small networks, we find that this improvement is of higher quality than all alternatives, even those with theoretical guarantees, such as unbiasedness, while being much cheaper to compute. We use this insight in reinforcement learning problems where small networks are used and demonstrate HesScale in second-order optimization and scaling the step-size parameter. In our experiments, HesScale optimizes faster than existing methods and improves stability through step-size scaling. These findings are promising for scaling second-order methods in larger models in the future.

Decisions made by machine learning models can have lasting impacts, making long-term fairness a critical consideration. It has been observed that ignoring the long-term effect and directly applying fairness criterion in static settings can actually worsen bias over time. To address biases in sequential decision-making, we introduce a long-term fairness concept named Equal Long-term Benefit Rate (ELBERT). This concept is seamlessly integrated into a Markov Decision Process (MDP) to consider the future effects of actions on long-term fairness, thus providing a unified framework for fair sequential decision-making problems. ELBERT effectively addresses the temporal discrimination issues found in previous long-term fairness notions. Additionally, we demonstrate that the policy gradient of Long-term Benefit Rate can be analytically simplified to standard policy gradients. This simplification makes conventional policy optimization methods viable for reducing bias, leading to our bias mitigation approach ELBERT-PO. Extensive experiments across various diverse sequential decision-making environments consistently reveal that ELBERT-PO significantly diminishes bias while maintaining high utility. Code is available at https://github.com/umd-huang-lab/ELBERT.

Identification and analysis of symmetrical patterns in the natural world have led to significant discoveries across various scientific fields, such as the formulation of gravitational laws in physics and advancements in the study of chemical structures. In this paper, we focus on exploiting Euclidean symmetries inherent in certain cooperative multi-agent reinforcement learning (MARL) problems and prevalent in many applications. We begin by formally characterizing a subclass of Markov games with a general notion of symmetries that admits the existence of symmetric optimal values and policies. Motivated by these properties, we design neural network architectures with symmetric constraints embedded as an inductive bias for multi-agent actor-critic methods. This inductive bias results in superior performance in various cooperative MARL benchmarks and impressive generalization capabilities such as zero-shot learning and transfer learning in unseen scenarios with repeated symmetric patterns.

In the rapidly changing healthcare landscape, the implementation of offline reinforcement learning (RL) in dynamic treatment regimes (DTRs) presents a mix of unprecedented opportunities and challenges. This position paper offers a critical examination of the current status of offline RL in the context of DTRs. We argue for a reassessment of applying RL in DTRs, citing concerns such as inconsistent and potentially inconclusive evaluation metrics, the absence of naive and supervised learning baselines, and the diverse choice of RL formulation in existing research. Through a case study with more than 17,000 evaluation experiments using a publicly available Sepsis dataset, we demonstrate that the performance of RL algorithms can significantly vary with changes in evaluation metrics and Markov Decision Process (MDP) formulations. Surprisingly, it is observed that in some instances, RL algorithms can be surpassed by random baselines subjected to policy evaluation methods and reward design. This calls for more careful policy evaluation and algorithm development in future DTR works. Additionally, we discussed potential enhancements toward more reliable development of RL-based dynamic treatment regimes and invited further discussion within the community. Code is available at https://github.com/GilesLuo/ReassessDTR.

We design differentially private regret-minimizing algorithms in the online convex optimization (OCO) framework. Unlike recent results, our algorithms and analyses do not require smoothness, thus yielding the first private regret bounds with an optimal leading-order term for non-smooth loss functions. Additionally, even for smooth losses, the resulting regret guarantees improve upon previous results in terms their dependence of dimension. Our results provide the best known rates for DP-OCO in all practical regimes of the privacy parameter, barring when it is exceptionally small. The principal innovation in our algorithm design is the use of sampling from strongly log-concave densities which satisfy the Log-Sobolev Inequality. The resulting concentration of measure allows us to obtain a better trade-off for the dimension factors than prior work, leading to improved results. Following previous works on DP-OCO, the proposed algorithm explicitly limits the number of switches via rejection sampling. Thus, independently of privacy constraints, the algorithm also provides improved results for online convex optimization with a switching budget.

Optimization and sampling based algorithms are two branches of methods in machine learning, while existing safe reinforcement learning (RL) algorithms are mainly based on optimization, it is still unclear whether sampling based methods can lead to desirable performance with safe policy. This paper formulates the Langevin policy for safe RL, and proposes Langevin Actor-Critic (LAC) to accelerate the process of policy inference. Concretely, instead of parametric policy, the proposed Langevin policy provides a stochastic process that directly infers actions, which is the numerical solver to the Langevin dynamic of actions on the continuous time. Furthermore, to make Langevin policy practical on RL tasks, the proposed LAC accumulates the transitions induced by Langevin policy and reproduces them with a generator. Finally, extensive empirical results show the effectiveness and superiority of LAC on the MuJoCo-based and Safety Gym tasks.

We propose a novel model-based reinforcement learning algorithm---Dynamics Learning and predictive control with Parameterized Actions (DLPA)---for Parameterized Action Markov Decision Processes (PAMDPs). The agent learns a parameterized-action-conditioned dynamics model and plans with a modified Model Predictive Path Integral control. We theoretically quantify the difference between the generated trajectory and the optimal trajectory during planning in terms of the value they achieved through the lens of Lipschitz Continuity. Our empirical results on several standard benchmarks show that our algorithm achieves superior sample efficiency and asymptotic performance than state-of-the-art PAMDP methods.

Temporal distances lie at the heart of many algorithms for planning, control, and reinforcement learning that involve reaching goals, allowing one to estimate the transit time between two states. However, prior attempts to define such temporal distances in stochastic settings have been stymied by an important limitation: these prior approaches do not satisfy the triangle inequality. This is not merely a definitional concern, but translates to an inability to generalize and find shortest paths. In this paper, we build on prior work in contrastive learning and quasimetrics to show how successor features learned by contrastive learning (after a change of variables) form a temporal distance that does satisfy the triangle inequality, even in stochastic settings. Importantly, this temporal distance is computationally efficient to estimate, even in high-dimensional and stochastic settings. Experiments in controlled settings and benchmark suites demonstrate that an RL algorithm based on these new temporal distances exhibits combinatorial generalization (i.e., "stitching") and can sometimes learn more quickly than prior methods, including those based on quasimetrics.

Conventional state representations in reinforcement learning often omit critical task-related details, presenting a significant challenge for value networks in establishing accurate mappings from states to task rewards. Traditional methods typically depend on extensive sample learning to enrich state representations with task-specific information, which leads to low sample efficiency and high time costs. Recently, surging knowledgeable large language models (LLM) have provided promising substitutes for prior injection with minimal human intervention. Motivated by this, we propose LLM-Empowered State Representation (LESR), a novel approach that utilizes LLM to autonomously generate task-related state representation codes which help to enhance the continuity of network mappings and facilitate efficient training. Experimental results demonstrate LESR exhibits high sample efficiency and outperforms state-of-the-art baselines by an average of **29%** in accumulated reward in Mujoco tasks and **30%** in success rates in Gym-Robotics tasks. Codes of LESR are accessible at https://github.com/thu-rllab/LESR.

In this paper, we propose an off-policy deep reinforcement learning (DRL) method utilizing the average reward criterion. While most existing DRL methods employ the discounted reward criterion, this can potentially lead to a discrepancy between the training objective and performance metrics in continuing tasks, making the average reward criterion a recommended alternative. We introduce RVI-SAC, an extension of the state-of-the-art off-policy DRL method, Soft Actor-Critic (SAC), to the average reward criterion. Our proposal consists of (1) Critic updates based on RVI Q-learning, (2) Actor updates introduced by the average reward soft policy improvement theorem, and (3) automatic adjustment of Reset Cost enabling the average reward reinforcement learning to be applied to tasks with termination. We apply our method to the Gymnasium's Mujoco tasks, a subset of locomotion tasks, and demonstrate that RVI-SAC shows competitive performance compared to existing methods.

Model-based reinforcement learning (MBRL) holds the promise of sample-efficient learning by utilizing a world model, which models how the environment works and typically encompasses components for two tasks: observation modeling and reward modeling. In this paper, through a dedicated empirical investigation, we gain a deeper understanding of the role each task plays in world models and uncover the overlooked potential of sample-efficient MBRL by mitigating the domination of either observation or reward modeling. Our key insight is that while prevalent approaches of explicit MBRL attempt to restore abundant details of the environment via observation models, it is difficult due to the environment's complexity and limited model capacity. On the other hand, reward models, while dominating implicit MBRL and adept at learning compact task-centric dynamics, are inadequate for sample-efficient learning without richer learning signals. Motivated by these insights and discoveries, we propose a simple yet effective approach, HarmonyDream, which automatically adjusts loss coefficients to maintain task harmonization, i.e. a dynamic equilibrium between the two tasks in world model learning. Our experiments show that the base MBRL method equipped with HarmonyDream gains 10%-69% absolute performance boosts on visual robotic tasks and sets a new state-of-the-art result on the Atari 100K benchmark. Code is available at https://github.com/thuml/HarmonyDream.

The popularity of multiagent reinforcement learning (MARL) is growing rapidly with the demand for real-world tasks that require swarm intelligence. However, a noticeable drawback of MARL is its low sample efficiency, which leads to a huge amount of interactions with the environment. Surprisingly, few MARL works focus on this practical problem especially in the parallel environment setting, which greatly hampers the application of MARL into the real world. In response to this gap, in this paper, we propose Multiagent Reinforcement Learning with Reset Replay (MARR) to greatly improve the sample efficiency of MARL by enabling MARL training at a high replay ratio in the parallel environment setting for the first time. To achieve this, first, a reset strategy is introduced for maintaining the network plasticity to ensure that MARL continually learns with a high replay ratio. Second, MARR incorporates a data augmentation technique to boost the sample efficiency further. Extensive experiments in SMAC and MPE show that MARR significantly improves the performance of various MARL approaches with much fewer environment interactions.

Model-based reinforcement learning agents utilizing transformers have shown improved sample efficiency due to their ability to model extended context, resulting in more accurate world models. However, for complex reasoning and planning tasks, these methods primarily rely on continuous representations. This complicates modeling of discrete properties of the real world such as disjoint object classes between which interpolation is not plausible. In this work, we introduce discrete abstract representations for transformer-based learning (DART), a sample-efficient method utilizing discrete representations for modeling both the world and learning behavior. We incorporate a transformer-decoder for auto-regressive world modeling and a transformer-encoder for learning behavior by attending to task-relevant cues in the discrete representation of the world model. For handling partial observability, we aggregate information from past time steps as memory tokens. DART outperforms previous state-of-the-art methods that do not use look-ahead search on the Atari 100k sample efficiency benchmark with a median human-normalized score of 0.790 and beats humans in 9 out of 26 games. We release our code at https://pranaval.github.io/DART/.

Foundation models such as GPT-4 are fine-tuned to avoid unsafe or otherwise problematic behavior, such as helping to commit crimes or producing racist text. One approach to fine-tuning, called reinforcement learning from human feedback, learns from humans’ expressed preferences over multiple outputs. Another approach is constitutional AI, in which the input from humans is a list of high-level principles. But how do we deal with potentially diverging input from humans? How can we aggregate the input into consistent data about “collective” preferences or otherwise use it to make collective choices about model behavior? In this paper, we argue that the field of social choice is well positioned to address these questions, and we discuss ways forward for this agenda, drawing on discussions in a recent workshop on Social Choice for AI Ethics and Safety held in Berkeley, CA, USA in December 2023.

Large Language Models (LLMs) are trained on massive text corpora, which are encoded with diverse personality traits. This triggers an interesting goal of eliciting a desired personality trait from the LLM, and probing its behavioral preferences. Accordingly, we formalize the persona elicitation task, aiming to customize LLM behaviors to align with a target persona. We present Persona In-Context Learning (PICLe), a novel persona elicitation framework grounded in Bayesian inference. At the core, PICLe introduces a new ICL example selection criterion based on likelihood ratio, which is designed to optimally guide the model in eliciting a specific target persona. We demonstrate the effectiveness of PICLe through extensive comparisons against baseline methods across three contemporary LLMs. Code is available at https://github.com/deeplearning-wisc/picle.