\Ac{MPC} and \ac{RL} are two powerful control strategies with, arguably, complementary advantages. In this work, we show how actor-critic \ac{RL} techniques can be leveraged to improve the performance of \ac{MPC}. The \ac{RL} critic is used as an approximation of the optimal value function, and an actor roll-out provides an initial guess for primal variables of the \ac{MPC}. A parallel control architecture is proposed where each \ac{MPC} instance is solved twice for different initial guesses. Besides the actor roll-out initialization, a shifted initialization from the previous solution is used. Thereafter, the actor and the critic are again used to approximately evaluate the infinite horizon cost of these trajectories. The control actions from the lowest-cost trajectory are applied to the system at each time step. We establish that the proposed algorithm is guaranteed to outperform the original \ac{RL} policy plus an error term that depends on the accuracy of the critic and decays with the horizon length of the \ac{MPC} formulation. Moreover, we do not require globally optimal solutions for these guarantees to hold. The approach is demonstrated on an illustrative toy example and an \ac{AD} overtaking scenario.
Safe Reinforcement Learning (SRL) aims to realize a safe learning process for Deep Reinforcement Learning (DRL) algorithms by incorporating safety constraints. However, the efficacy of SRL approaches often relies on accurate function approximations, which are notably challenging to achieve in the early learning stages due to data insufficiency. To address this issue, we introduce in this work a novel Generalizable Safety enhancer (GenSafe) that is able to overcome the challenge of data insufficiency and enhance the performance of SRL approaches. Leveraging model order reduction techniques, we first propose an innovative method to construct a Reduced Order Markov Decision Process (ROMDP) as a low-dimensional approximator of the original safety constraints. Then, by solving the reformulated ROMDP-based constraints, GenSafe refines the actions of the agent to increase the possibility of constraint satisfaction. Essentially, GenSafe acts as an additional safety layer for SRL algorithms. We evaluate GenSafe on multiple SRL approaches and benchmark problems. The results demonstrate its capability to improve safety performance, especially in the early learning phases, while maintaining satisfactory task performance. Our proposed GenSafe not only offers a novel measure to augment existing SRL methods but also shows broad compatibility with various SRL algorithms, making it applicable to a wide range of systems and SRL problems.
This study considers the attack on reinforcement learning agents where the adversary aims to control the victim's behavior as specified by the adversary by adding adversarial modifications to the victim's state observation. While some attack methods reported success in manipulating the victim agent's behavior, these methods often rely on environment-specific heuristics. In addition, all existing attack methods require white-box access to the victim's policy. In this study, we propose a novel method for manipulating the victim agent in the black-box (i.e., the adversary is allowed to observe the victim's state and action only) and no-box (i.e., the adversary is allowed to observe the victim's state only) setting without requiring environment-specific heuristics. Our attack method is formulated as a bi-level optimization problem that is reduced to a distribution matching problem and can be solved by an existing imitation learning algorithm in the black-box and no-box settings. Empirical evaluations on several reinforcement learning benchmarks show that our proposed method has superior attack performance to baselines.
We present a novel inductive generalization framework for RL from logical specifications. Many interesting tasks in RL environments have a natural inductive structure. These inductive tasks have similar overarching goals but they differ inductively in low-level predicates and distributions. We present a generalization procedure that leverages this inductive relationship to learn a higher-order function, a policy generator, that generates appropriately adapted policies for instances of an inductive task in a zero-shot manner. An evaluation of the proposed approach on a set of challenging control benchmarks demonstrates the promise of our framework in generalizing to unseen policies for long-horizon tasks.
Quantum machine learning models use encoding circuits to map data into a quantum Hilbert space. While it is well known that the architecture of these circuits significantly influences core properties of the resulting model, they are often chosen heuristically. In this work, we present a novel approach using reinforcement learning techniques to generate problem-specific encoding circuits to improve the performance of quantum machine learning models. By specifically using a model-based reinforcement learning algorithm, we reduce the number of necessary circuit evaluations during the search, providing a sample-efficient framework. In contrast to previous search algorithms, our method uses a layered circuit structure that significantly reduces the search space. Additionally, our approach can account for multiple objectives such as solution quality, hardware restrictions and circuit depth. We benchmark our tailored circuits against various reference models, including models with problem-agnostic circuits and classical models. Our results highlight the effectiveness of problem-specific encoding circuits in enhancing QML model performance.
Nowadays, a significant share of the Business-to-Consumer sector is based on online platforms like Amazon and Alibaba and uses Artificial Intelligence for pricing strategies. This has sparked debate on whether pricing algorithms may tacitly collude to set supra-competitive prices without being explicitly designed to do so. Our study addresses these concerns by examining the risk of collusion when Reinforcement Learning algorithms are used to decide on pricing strategies in competitive markets. Prior research in this field focused on Tabular Q-learning (TQL) and led to opposing views on whether learning-based algorithms can lead to supra-competitive prices. Our work contributes to this ongoing discussion by providing a more nuanced numerical study that goes beyond TQL by additionally capturing off- and on-policy Deep Reinforcement Learning (DRL) algorithms. We study multiple Bertrand oligopoly variants and show that algorithmic collusion depends on the algorithm used. In our experiments, TQL exhibits higher collusion and price dispersion phenomena compared to DRL algorithms. We show that the severity of collusion depends not only on the algorithm used but also on the characteristics of the market environment. We further find that Proximal Policy Optimization appears to be less sensitive to collusive outcomes compared to other state-of-the-art DRL algorithms.
In this work, we use optimal control to change the behavior of a deep reinforcement learning policy by optimizing directly in the policy's latent space. We hypothesize that distinct behavioral patterns, termed behavioral modes, can be identified within certain regions of a deep reinforcement learning policy's latent space, meaning that specific actions or strategies are preferred within these regions. We identify these behavioral modes using latent space dimension-reduction with \ac*{pacmap}. Using the actions generated by the optimal control procedure, we move the system from one behavioral mode to another. We subsequently utilize these actions as a filter for interpreting the neural network policy. The results show that this approach can impose desired behavioral modes in the policy, demonstrated by showing how a failed episode can be made successful and vice versa using the lunar lander reinforcement learning environment.
Learning representations for reinforcement learning (RL) has shown much promise for continuous control. We propose an efficient representation learning method using only a self-supervised latent-state consistency loss. Our approach employs an encoder and a dynamics model to map observations to latent states and predict future latent states, respectively. We achieve high performance and prevent representation collapse by quantizing the latent representation such that the rank of the representation is empirically preserved. Our method, named iQRL: implicitly Quantized Reinforcement Learning, is straightforward, compatible with any model-free RL algorithm, and demonstrates excellent performance by outperforming other recently proposed representation learning methods in continuous control benchmarks from DeepMind Control Suite.
The growing complexity of power system management has led to an increased interest in reinforcement learning (RL). To validate their effectiveness, RL algorithms have to be evaluated across multiple case studies. Case study design is an arduous task requiring the consideration of many aspects, among them the influence of available forecasts and the level of decentralization in the control structure. Furthermore, vanilla RL controllers cannot themselves ensure the satisfaction of system constraints, which makes devising a safeguarding mechanism a necessary task for every case study before deploying the system. To address these shortcomings, we introduce the Python tool CommonPower, the first general framework for the modeling and simulation of power system management tailored towards machine learning. Its modular architecture enables users to focus on specific elements without having to implement a simulation environment. Another unique contribution of CommonPower is the automatic synthesis of model predictive controllers and safeguards. Beyond offering a unified interface for single-agent RL, multi-agent RL, and optimal control, CommonPower includes a training pipeline for machine-learning-based forecasters as well as a flexible mechanism for incorporating feedback of safeguards into the learning updates of RL controllers.
Sample efficiency remains a key challenge in multi-agent reinforcement learning (MARL). A promising approach is to learn a meaningful latent representation space through auxiliary learning objectives alongside the MARL objective to aid in learning a successful control policy. In our work, we present MAPO-LSO (Multi-Agent Policy Optimization with Latent Space Optimization) which applies a form of comprehensive representation learning devised to supplement MARL training. Specifically, MAPO-LSO proposes a multi-agent extension of transition dynamics reconstruction and self-predictive learning that constructs a latent state optimization scheme that can be trivially extended to current state-of-the-art MARL algorithms. Empirical results demonstrate MAPO-LSO to show notable improvements in sample efficiency and learning performance compared to its vanilla MARL counterpart without any additional MARL hyperparameter tuning on a diverse suite of MARL tasks.
Reinforcement learning from human feedback (RLHF) is a prevalent approach to align AI systems with human values by learning rewards from human preference data. Due to various reasons, however, such data typically takes the form of rankings over pairs of trajectory segments, which fails to capture the varying strengths of preferences across different pairs. In this paper, we propose a novel adaptive preference loss, underpinned by distributionally robust optimization (DRO), designed to address this uncertainty in preference strength. By incorporating an adaptive scaling parameter into the loss for each pair, our method increases the flexibility of the reward function. Specifically, it assigns small scaling parameters to pairs with ambiguous preferences, leading to more comparable rewards, and large scaling parameters to those with clear preferences for more distinct rewards. Computationally, our proposed loss function is strictly convex and univariate with respect to each scaling parameter, enabling its efficient optimization through a simple second-order algorithm. Our method is versatile and can be readily adapted to various preference optimization frameworks, including direct preference optimization (DPO). Our experiments with robotic control and natural language generation with large language models (LLMs) show that our method not only improves policy performance but also aligns reward function selection more closely with policy optimization, simplifying the hyperparameter tuning process.
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.
Reinforcement learning-based recommender systems have recently gained popularity. However, due to the typical limitations of simulation environments (e.g., data inefficiency), most of the work cannot be broadly applied in all domains. To counter these challenges, recent advancements have leveraged offline reinforcement learning methods, notable for their data-driven approach utilizing offline datasets. A prominent example of this is the Decision Transformer. Despite its popularity, the Decision Transformer approach has inherent drawbacks, particularly evident in recommendation methods based on it. This paper identifies two key shortcomings in existing Decision Transformer-based methods: a lack of stitching capability and limited effectiveness in online adoption. In response, we introduce a novel methodology named Max-Entropy enhanced Decision Transformer with Reward Relabeling for Offline RLRS (EDT4Rec). Our approach begins with a max entropy perspective, leading to the development of a max entropy enhanced exploration strategy. This strategy is designed to facilitate more effective exploration in online environments. Additionally, to augment the model's capability to stitch sub-optimal trajectories, we incorporate a unique reward relabeling technique. To validate the effectiveness and superiority of EDT4Rec, we have conducted comprehensive experiments across six real-world offline datasets and in an online simulator.
Reinforcement Learning from Human Feedback (RLHF) has enabled significant advancements within language modeling for powerful, instruction-following models. However, the alignment of these models remains a pressing challenge as the policy tends to overfit the learned ``proxy" reward model past an inflection point of utility as measured by a ``gold" reward model that is more performant -- a phenomenon known as overoptimisation. Prior work has mitigated this issue by computing a pessimistic statistic over an ensemble of reward models, which is common in Offline Reinforcement Learning but incredibly costly for language models with high memory requirements, making such approaches infeasible for sufficiently large models. To this end, we propose using a shared encoder but separate linear heads. We find this leads to similar performance as the full ensemble while allowing tremendous savings in memory and time required for training for models of similar size.
The evaluation of summary quality encompasses diverse dimensions such as consistency, coherence, relevance, and fluency. However, existing summarization methods often target a specific dimension, facing challenges in generating well-balanced summaries across multiple dimensions. In this paper, we propose multi-objective reinforcement learning tailored to generate balanced summaries across all four dimensions. We introduce two multi-dimensional optimization (MDO) strategies for adaptive learning: 1) MDO_min, rewarding the current lowest dimension score, and 2) MDO_pro, optimizing multiple dimensions similar to multi-task learning, resolves conflicting gradients across dimensions through gradient projection. Unlike prior ROUGE-based rewards relying on reference summaries, we use a QA-based reward model that aligns with human preferences. Further, we discover the capability to regulate the length of summaries by adjusting the discount factor, seeking the generation of concise yet informative summaries that encapsulate crucial points. Our approach achieved substantial performance gains compared to baseline models on representative summarization datasets, particularly in the overlooked dimensions.
Meta reinforcement learning sets a distribution over a set of tasks on which the agent can train at will, then is asked to learn an optimal policy for any test task efficiently. In this paper, we consider a finite set of tasks modeled through Markov decision processes with various dynamics. We assume to have endured a long training phase, from which the set of tasks is perfectly recovered, and we focus on regret minimization against the optimal policy in the unknown test task. Under a separation condition that states the existence of a state-action pair revealing a task against another, Chen et al. (2022) show that $O(M^2 \log(H))$ regret can be achieved, where $M, H$ are the number of tasks in the set and test episodes, respectively. In our first contribution, we demonstrate that the latter rate is nearly optimal by developing a novel lower bound for test-time regret minimization under separation, showing that a linear dependence with $M$ is unavoidable. Then, we present a family of stronger yet reasonable assumptions beyond separation, which we call strong identifiability, enabling algorithms achieving fast rates $\log (H)$ and sublinear dependence with $M$ simultaneously. Our paper provides a new understanding of the statistical barriers of test-time regret minimization and when fast rates can be achieved.
We study reinforcement learning (RL) problems in which agents observe the reward or transition realizations at their current state before deciding which action to take. Such observations are available in many applications, including transactions, navigation and more. When the environment is known, previous work shows that this lookahead information can drastically increase the collected reward. However, outside of specific applications, existing approaches for interacting with unknown environments are not well-adapted to these observations. In this work, we close this gap and design provably-efficient learning algorithms able to incorporate lookahead information. To achieve this, we perform planning using the empirical distribution of the reward and transition observations, in contrast to vanilla approaches that only rely on estimated expectations. We prove that our algorithms achieve tight regret versus a baseline that also has access to lookahead information - linearly increasing the amount of collected reward compared to agents that cannot handle lookahead information.
The Generative Flow Network (GFlowNet) is a probabilistic framework in which an agent learns a stochastic policy and flow functions to sample objects with probability proportional to an unnormalized reward function. GFlowNets share a strong resemblance to reinforcement learning (RL), that typically aims to maximize reward, due to their sequential decision-making processes. Recent works have studied connections between GFlowNets and maximum entropy (MaxEnt) RL, which modifies the standard objective of RL agents by learning an entropy-regularized objective. However, a critical theoretical gap persists: despite the apparent similarities in their sequential decision-making nature, a direct link between GFlowNets and standard RL has yet to be discovered, while bridging this gap could further unlock the potential of both fields. In this paper, we establish a new connection between GFlowNets and policy evaluation for a uniform policy. Surprisingly, we find that the resulting value function for the uniform policy has a close relationship to the flows in GFlowNets. Leveraging these insights, we further propose a novel rectified policy evaluation (RPE) algorithm, which achieves the same reward-matching effect as GFlowNets, offering a new perspective. We compare RPE, MaxEnt RL, and GFlowNets in a number of benchmarks, and show that RPE achieves competitive results compared to previous approaches. This work sheds light on the previously unexplored connection between (non-MaxEnt) RL and GFlowNets, potentially opening new avenues for future research in both fields.
Traditional reinforcement learning (RL) methods typically employ a fixed control loop, where each cycle corresponds to an action. This rigidity poses challenges in practical applications, as the optimal control frequency is task-dependent. A suboptimal choice can lead to high computational demands and reduced exploration efficiency. Variable Time Step Reinforcement Learning (VTS-RL) addresses these issues by using adaptive frequencies for the control loop, executing actions only when necessary. This approach, rooted in reactive programming principles, reduces computational load and extends the action space by including action durations. However, VTS-RL's implementation is often complicated by the need to tune multiple hyperparameters that govern exploration in the multi-objective action-duration space (i.e., balancing task performance and number of time steps to achieve a goal). To overcome these challenges, we introduce the Multi-Objective Soft Elastic Actor-Critic (MOSEAC) method. This method features an adaptive reward scheme that adjusts hyperparameters based on observed trends in task rewards during training. This scheme reduces the complexity of hyperparameter tuning, requiring a single hyperparameter to guide exploration, thereby simplifying the learning process and lowering deployment costs. We validate the MOSEAC method through simulations in a Newtonian kinematics environment, demonstrating high task and training performance with fewer time steps, ultimately lowering energy consumption. This validation shows that MOSEAC streamlines RL algorithm deployment by automatically tuning the agent control loop frequency using a single parameter. Its principles can be applied to enhance any RL algorithm, making it a versatile solution for various applications.
Powered by massive curated training data, Segment Anything Model (SAM) has demonstrated its impressive generalization capabilities in open-world scenarios with the guidance of prompts. However, the vanilla SAM is class agnostic and heavily relies on user-provided prompts to segment objects of interest. Adapting this method to diverse tasks is crucial for accurate target identification and to avoid suboptimal segmentation results. In this paper, we propose a novel framework, termed AlignSAM, designed for automatic prompting for aligning SAM to an open context through reinforcement learning. Anchored by an agent, AlignSAM enables the generality of the SAM model across diverse downstream tasks while keeping its parameters frozen. Specifically, AlignSAM initiates a prompting agent to iteratively refine segmentation predictions by interacting with the foundational model. It integrates a reinforcement learning policy network to provide informative prompts to the foundational models. Additionally, a semantic recalibration module is introduced to provide fine-grained labels of prompts, enhancing the model's proficiency in handling tasks encompassing explicit and implicit semantics. Experiments conducted on various challenging segmentation tasks among existing foundation models demonstrate the superiority of the proposed AlignSAM over state-of-the-art approaches. Project page: \url{https://github.com/Duojun-Huang/AlignSAM-CVPR2024}.