Safe Reinforcement Learning (SRL) algorithms aim to learn a policy that maximizes the reward while satisfying the safety constraints. One of the challenges in SRL is that it is often difficult to balance the two objectives of reward maximization and safety constraint satisfaction. Existing algorithms utilize constraint optimization techniques like penalty-based, barrier penalty-based, and Lagrangian-based dual or primal policy optimizations methods. However, they suffer from training oscillations and approximation errors, which impact the overall learning objectives. This paper proposes the Permeable Penalty Barrier-based Policy Optimization (P2BPO) algorithm that addresses this issue by allowing a small fraction of penalty beyond the penalty barrier, and a parameter is used to control this permeability. In addition, an adaptive penalty parameter is used instead of a constant one, which is initialized with a low value and increased gradually as the agent violates the safety constraints. We have also provided a theoretical proof of the proposed method's performance guarantee bound, which ensures that P2BPO can learn a policy satisfying the safety constraints with high probability while achieving a higher expected reward. Furthermore, we compare P2BPO with other SRL algorithms on various SRL tasks and demonstrate that it achieves better rewards while adhering to the constraints.
论文检索
输入标题、作者或关键词，从 2,866 篇学术成果中精准定位
Tree ensembles are one of the most widely used model classes. However, these models are susceptible to adversarial examples, which are slightly perturbed examples that elicit a misprediction. There has been significant research on designing approaches to verify the robustness of tree ensembles to such attacks. However, existing verification algorithms for tree ensembles are only able to analyze binary classifiers and hence address multiclass problems by reducing them to binary ones using a one-versus-other strategy. In this paper, we show that naively applying this strategy can yield incorrect results in certain situations. We address this shortcoming by proposing a novel approximate heuristic approach to verification for multiclass tree ensembles. Our approach is based on a novel generalization of the verification task, which we show emits other relevant verification queries.
Neural-based, data-driven analysis and control of dynamical systems have been recently investigated and have shown great promise, e.g. for safety verification or stability analysis. Indeed, not only do neural networks allow for an entirely model-free, data-driven approach, but also for handling arbitrary complex functions via their power of representation (as opposed to, e.g. algebraic optimization techniques that are restricted to polynomial functions). Whilst classical Lyapunov techniques allow to provide a formal and robust guarantee of stability of a switched dynamical system, very little is yet known about correctness guarantees for Neural Lyapunov functions, nor about their performance (amount of data needed for a certain accuracy). We formally introduce Neural Lyapunov functions for the stability analysis of switched linear systems: we benchmark them on this paradigmatic problem, which is notoriously difficult (and in general Turing-undecidable), but which admits existing recently-developed technologies and theoretical results. Inspired by switched systems theory, we provide theoretical guarantees on the representative power of neural networks, leveraging recent results from the ML community. We additionally experimentally display how Neural Lyapunov functions compete with state-of-the-art results and techniques, while admitting a wide range of improvement, both in theory and in practice. This study intends to improve our understanding of the opportunities and current limitations of neural-based data-driven analysis and control of complex dynamical systems.
Deep neural networks are characterized by multiple symmetrical, equi-loss solutions that are redundant. Thus, the order of neurons in a layer and feature maps can be given arbitrary permutations, without affecting (or minimally affecting) their output. If we shuffle these neurons, or if we apply to them some perturbations (like fine-tuning) can we put them back in the original order i.e. re-synchronize? Is there a possible corruption threat? Answering these questions is important for applications like neural network white-box watermarking for ownership tracking and integrity verification. We advance a method to re-synchronize the order of permuted neurons. Our method is also effective if neurons are further altered by parameter pruning, quantization, and fine-tuning, showing robustness to integrity attacks. Additionally, we provide theoretical and practical evidence for the usual means to corrupt the integrity of the model, resulting in a solution to counter it. We test our approach on popular computer vision datasets and models, and we illustrate the threat and our countermeasure on a popular white-box watermarking method.
TTTS: Tree Test Time Simulation for Enhancing Decision Tree Robustness against Adversarial Examples
PDF ↗Decision trees are widely used for addressing learning tasks involving tabular data. Yet, they are susceptible to adversarial attacks. In this paper, we present Tree Test Time Simulation (TTTS), a novel inference-time methodology that incorporates Monte Carlo simulations into decision trees to enhance their robustness. TTTS introduces a probabilistic modification to the decision path, without altering the underlying tree structure. Our comprehensive empirical analysis of 50 datasets yields promising results. Without the presence of any attacks, TTTS has successfully improved model performance from an AUC of 0.714 to 0.773. Under the challenging conditions of white-box attacks, TTTS demonstrated its robustness by boosting performance from an AUC of 0.337 to 0.680. Even when subjected to black-box attacks, TTTS maintains high accuracy and enhances the model's performance from an AUC of 0.628 to 0.719. Compared to defenses such as Feature Squeezing, TTTS proves to be much more effective. We also found that TTTS exhibits similar robustness in decision forest settings across different attacks.
Conformal prediction is a statistical tool for producing prediction regions of machine learning models that are valid with high probability. However, applying conformal prediction to time series data leads to conservative prediction regions. In fact, to obtain prediction regions over T time steps with confidence 1--delta, previous works require that each individual prediction region is valid with confidence 1--delta/T. We propose an optimization-based method for reducing this conservatism to enable long horizon planning and verification when using learning-enabled time series predictors. Instead of considering prediction errors individually at each time step, we consider a parameterized prediction error over multiple time steps. By optimizing the parameters over an additional dataset, we find prediction regions that are not conservative. We show that this problem can be cast as a mixed integer linear complementarity program (MILCP), which we then relax into a linear complementarity program (LCP). Additionally, we prove that the relaxed LP has the same optimal cost as the original MILCP. Finally, we demonstrate the efficacy of our method on case studies using pedestrian trajectory predictors and F16 fighter jet altitude predictors.
Constrained Meta-Reinforcement Learning for Adaptable Safety Guarantee with Differentiable Convex Programming
PDF ↗Despite remarkable achievements in artificial intelligence, the deployability of learning-enabled systems in high-stakes real-world environments still faces persistent challenges. For example, in safety-critical domains like autonomous driving, robotic manipulation, and healthcare, it is crucial not only to achieve high performance but also to comply with given constraints. Furthermore, adaptability becomes paramount in non-stationary domains, where environmental parameters are subject to change. While safety and adaptability are recognized as key qualities for the new generation of AI, current approaches have not demonstrated effective adaptable performance in constrained settings. Hence, this paper breaks new ground by studying the unique challenges of ensuring safety in nonstationary environments by solving constrained problems through the lens of the meta-learning approach (learning to learn). While unconstrained meta-learning already encounters complexities in end to end differentiation of the loss due to the bi-level nature, its constrained counterpart introduces an additional layer of difficulty, since the constraints imposed on task-level updates complicate the differentiation process. To address the issue, we first employ successive convex-constrained policy updates across multiple tasks with differentiable convex programming, which allows meta-learning in constrained scenarios by enabling end-to-end differentiation. This approach empowers the agent to rapidly adapt to new tasks under nonstationarity while ensuring compliance with safety constraints. We also provide a theoretical analysis demonstrating guaranteed monotonic improvement of our approach, justifying our algorithmic designs. Extensive simulations across diverse environments provide empirical validation with significant improvement over established benchmarks.
Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning
PDF ↗Despite the great success of large language models (LLMs) in various tasks, they suffer from generating hallucinations. We introduce Truth Forest, a method that enhances truthfulness in LLMs by uncovering hidden truth representations using multi-dimensional orthogonal probes. Specifically, it creates multiple orthogonal bases for modeling truth by incorporating orthogonal constraints into the probes. Moreover, we introduce Random Peek, a systematic technique considering an extended range of positions within the sequence, reducing the gap between discerning and generating truth features in LLMs. By employing this approach, we improved the truthfulness of Llama-2-7B from 40.8% to 74.5% on TruthfulQA. Likewise, significant improvements are observed in fine-tuned models. We conducted a thorough analysis of truth features using probes. Our visualization results show that orthogonal probes capture complementary truth-related features, forming well-defined clusters that reveal the inherent structure of the dataset.
Rethinking the Development of Large Language Models from the Causal Perspective: A Legal Text Prediction Case Study
PDF ↗While large language models (LLMs) exhibit impressive performance on a wide range of NLP tasks, most of them fail to learn the causality from correlation, which disables them from learning rationales for predicting. Rethinking the whole developing process of LLMs is of great urgency as they are adopted in various critical tasks that need rationales, including legal text prediction (e.g., legal judgment prediction). In this paper, we first explain the underlying theoretical mechanism of their failure and argue that both the data imbalance and the omission of causality in model design and selection render the current training-testing paradigm failed to select the unique causality-based model from correlation-based models. Second, we take the legal text prediction task as the testbed and reconstruct the developing process of LLMs by simultaneously infusing causality into model architectures and organizing causality-based adversarial attacks for evaluation. Specifically, we base our reconstruction on our theoretical analysis and propose a causality-aware self-attention mechanism (CASAM), which prevents LLMs from entangling causal and non-causal information by restricting the interaction between causal and non-causal words. Meanwhile, we propose eight kinds of legal-specific attacks to form causality-based model selection. Our extensive experimental results demonstrate that our proposed CASAM achieves state-of-the-art (SOTA) performances and the strongest robustness on three commonly used legal text prediction benchmarks. We make our code publicly available at https://github.com/Carrot-Red/Rethink-LLM-development.
CCTR: Calibrating Trajectory Prediction for Uncertainty-Aware Motion Planning in Autonomous Driving
PDF ↗Autonomous driving systems rely on precise trajectory prediction for safe and efficient motion planning. Despite considerable efforts to enhance prediction accuracy, inherent uncertainties persist due to data noise and incomplete observations. Many strategies entail formalizing prediction outcomes into distributions and utilizing variance to represent uncertainty. However, our experimental investigation reveals that existing trajectory prediction models yield unreliable uncertainty estimates, necessitating additional customized calibration processes. On the other hand, directly applying current calibration techniques to prediction outputs may yield sub-optimal results due to using a universal scaler for all predictions and neglecting informative data cues. In this paper, we propose Customized Calibration Temperature with Regularizer (CCTR), a generic framework that calibrates the output distribution. Specifically, CCTR 1) employs a calibration-based regularizer to align output variance with the discrepancy between prediction and ground truth and 2) generates a tailor-made temperature scaler for each prediction using a post-processing network guided by context and historical information. Extensive evaluation involving multiple prediction and planning methods demonstrates the superiority of CCTR over existing calibration algorithms and uncertainty-aware methods, with significant improvements of 11%-22% in calibration quality and 17%-46% in motion planning.
Deploying deep learning models in safety-critical applications remains a very challenging task, mandating the provision of assurances for the dependable operation of these models. Uncertainty quantification (UQ) methods estimate the model’s confidence per prediction, informing decision-making by considering the effect of randomness and model misspecification. Despite the advances of state-of-the-art UQ methods, they are computationally expensive or produce conservative prediction sets/intervals. We introduce MC-CP, a novel hybrid UQ method that combines a new adaptive Monte Carlo (MC) dropout method with conformal prediction (CP). MC-CP adaptively modulates the traditional MC dropout at runtime to save memory and computation resources, enabling predictions to be consumed by CP, yielding robust prediction sets/intervals. Throughout comprehensive experiments, we show that MC-CP delivers significant improvements over comparable UQ methods, like MC dropout, RAPS and CQR, both in classification and regression benchmarks. MC-CP can be easily added to existing models, making its deployment simple. The MC-CP code and replication package is available at https://github.com/team-daniel/MC-CP.
Recent advances in learning techniques have garnered attention for their applicability to a diverse range of real-world sequential decision-making problems. Yet, many practical applications have critical constraints for operation in real environments. Most learning solutions often neglect the risk of failing to meet these constraints, hindering their implementation in real-world contexts. In this paper, we propose a risk-aware decision-making framework for contextual bandit problems, accommodating constraints and continuous action spaces. Our approach employs an actor multi-critic architecture, with each critic characterizing the distribution of performance and constraint metrics. Our framework is designed to cater to various risk levels, effectively balancing constraint satisfaction against performance. To demonstrate the effectiveness of our approach, we first compare it against state-of-the-art baseline methods in a synthetic environment, highlighting the impact of intrinsic environmental noise across different risk configurations. Finally, we evaluate our framework in a real-world use case involving a 5G mobile network where only our approach satisfies consistently the system constraint (a signal processing reliability target) with a small performance toll (8.5% increase in power consumption).
On the Importance of Application-Grounded Experimental Design for Evaluating Explainable ML Methods
PDF ↗Most existing evaluations of explainable machine learning (ML) methods rely on simplifying assumptions or proxies that do not reflect real-world use cases; the handful of more robust evaluations on real-world settings have shortcomings in their design, generally leading to overestimation of methods' real-world utility. In this work, we seek to address this by conducting a study that evaluates post-hoc explainable ML methods in a setting consistent with the application context and provide a template for future evaluation studies. We modify and improve a prior study on e-commerce fraud detection by relaxing the original work's simplifying assumptions that departed from the deployment context. Our study finds no evidence for the utility of the tested explainable ML methods in the context, which is a drastically different conclusion from the earlier work. This highlights how seemingly trivial experimental design choices can yield misleading conclusions about method utility. In addition, our work carries lessons about the necessity of not only evaluating explainable ML methods using tasks, data, users, and metrics grounded in the intended application context but also developing methods tailored to specific applications, moving beyond general-purpose explainable ML methods.
Advancements in mathematical programming have made it possible to efficiently tackle large-scale real-world problems that were deemed intractable just a few decades ago. However, provably optimal solutions may not be accepted due to the perception of optimization software as a black box. Although well understood by scientists, this lacks easy accessibility for practitioners. Hence, we advocate for introducing the explainability of a solution as another evaluation criterion, next to its objective value, which enables us to find trade-off solutions between these two criteria. Explainability is attained by comparing against (not necessarily optimal) solutions that were implemented in similar situations in the past. Thus, solutions are preferred that exhibit similar features. Although we prove that already in simple cases the explainable model is NP-hard, we characterize relevant polynomially solvable cases such as the explainable shortest path problem. Our numerical experiments on both artificial as well as real-world road networks show the resulting Pareto front. It turns out that the cost of enforcing explainability can be very small.
Most pre-trained learning systems are known to suffer from bias, which typically emerges from the data, the model, or both. Measuring and quantifying bias and its sources is a challenging task and has been extensively studied in image captioning. Despite the significant effort in this direction, we observed that existing metrics lack consistency in the inclusion of the visual signal. In this paper, we introduce a new bias assessment metric, dubbed ImageCaptioner2, for image captioning. Instead of measuring the absolute bias in the model or the data, ImageCaptioner2pay more attention to the bias introduced by the model w.r.t the data bias, termed bias amplification. Unlike the existing methods, which only evaluate the image captioning algorithms based on the generated captions only, ImageCaptioner2incorporates the image while measuring the bias. In addition, we design a formulation for measuring the bias of generated captions as prompt-based image captioning instead of using language classifiers. Finally, we apply our ImageCaptioner2metric across 11 different image captioning architectures on three different datasets, i.e., MS-COCO caption dataset, Artemis V1, and Artemis V2, and on three different protected attributes, i.e., gender, race, and emotions. Consequently, we verify the effectiveness of our ImageCaptioner2metric by proposing Anonymous-Bench, which is a novel human evaluation paradigm for bias metrics. Our metric shows significant superiority over the recent bias metric; LIC, in terms of human alignment, where the correlation scores are 80% and 54% for our metric and LIC, respectively. The code and more details are available at https://eslambakr.github.io/imagecaptioner2.github.io/.
The integration of Evolutionary Algorithm (EA) and Reinforcement Learning (RL) has emerged as a promising approach for tackling some challenges in RL, such as sparse rewards, lack of exploration, and brittle convergence properties. However, existing methods often employ actor networks as individuals of EA, which may constrain their exploratory capabilities, as the entire actor population will stop evolution when the critic network in RL falls into local optimal. To alleviate this issue, this paper introduces a Two-stage Evolutionary Reinforcement Learning (TERL) framework that maintains a population containing both actor and critic networks. TERL divides the learning process into two stages. In the initial stage, individuals independently learn actor-critic networks, which are optimized alternatively by RL and Particle Swarm Optimization (PSO). This dual optimization fosters greater exploration, curbing susceptibility to local optima. Shared information from a common replay buffer and PSO algorithm substantially mitigates the computational load of training multiple agents. In the subsequent stage, TERL shifts to a refined exploitation phase. Here, only the best individual undergoes further refinement, while the rest individuals continue PSO-based optimization. This allocates more computational resources to the best individual for yielding superior performance. Empirical assessments, conducted across a range of continuous control problems, validate the efficacy of the proposed TERL paradigm.
The Metropolis algorithm can cope with local optima by accepting inferior solutions with suitably small probability. That this can work well was not only observed in empirical research, but also via mathematical runtime analyses on single-objective benchmarks. This paper takes several steps towards understanding, again via theoretical means, whether such advantages can also be obtained in multi-objective optimization. The original Metropolis algorithm has two components, one-bit mutation and the acceptance strategy, which allows accepting inferior solutions. When adjusting the acceptance strategy to multi-objective optimization in the way that an inferior solution that is accepted replaces its parent, then the Metropolis algorithm is not very efficient on our multi-objective version of the multimodal DLB benchmark called DLTB. With one-bit mutation, this multi-objective Metropolis algorithm cannot optimize the DLTB problem, with standard bit-wise mutation it needs at least Ω(n^5) time to cover the full Pareto front. In contrast, we show that many other multi-objective optimizers, namely the GSEMO, SMS-EMOA, and NSGA-II, only need time O(n^4). When keeping the parent when an inferior point is accepted, the multi-objective Metropolis algorithm both with one-bit or standard bit-wise mutation solves the DLTB problem efficiently, with one-bit mutation experimentally leading to better results than several other algorithms. Overall, our work suggests that the general mechanism of the Metropolis algorithm can be interesting in multi-objective optimization, but that the implementation details can have a huge impact on the performance.
The widely used multiobjective optimizer NSGA-II was recently proven to have considerable difficulties in many-objective optimization. In contrast, experimental results in the literature show a good performance of the SMS-EMOA, which can be seen as a steady-state NSGA-II that uses the hypervolume contribution instead of the crowding distance as the second selection criterion. This paper conducts the first rigorous runtime analysis of the SMS-EMOA for many-objective optimization. To this aim, we first propose a many-objective counterpart, the m-objective mOJZJ problem, of the bi-objective OJZJ benchmark, which is the first many-objective multimodal benchmark used in a mathematical runtime analysis. We prove that SMS-EMOA computes the full Pareto front of this benchmark in an expected number of O(M^2 n^k) iterations, where n denotes the problem size (length of the bit-string representation), k the gap size (a difficulty parameter of the problem), and M=(2n/m-2k+3)^(m/2) the size of the Pareto front. This result together with the existing negative result on the original NSGA-II shows that in principle, the general approach of the NSGA-II is suitable for many-objective optimization, but the crowding distance as tie-breaker has deficiencies. We obtain three additional insights on the SMS-EMOA. Different from a recent result for the bi-objective OJZJ benchmark, the stochastic population update often does not help for mOJZJ. It results in a 1/Θ(min(Mk^(1/2)/2^(k/2),1)) speed-up, which is Θ(1) for large m such as m>k. On the positive side, we prove that heavy-tailed mutation still results in a speed-up of order k^(0.5+k-β). Finally, we conduct the first runtime analyses of the SMS-EMOA on the bi-objective OneMinMax and LOTZ benchmarks and show that it has a performance comparable to the GSEMO and the NSGA-II.
Jointly Improving the Sample and Communication Complexities in Decentralized Stochastic Minimax Optimization
PDF ↗We propose a novel single-loop decentralized algorithm, DGDA-VR, for solving the stochastic nonconvex strongly-concave minimax problems over a connected network of agents, which are equipped with stochastic first-order oracles to estimate their local gradients. DGDA-VR, incorporating variance reduction, achieves O(ε^−3) oracle complexity and O(ε^−2) communication complexity without resorting to multi-communication rounds – both are optimal, i.e., matching the lower bounds for this class of problems. Since DGDA-VR does not require multiple communication rounds, it is applicable to a broader range of decentralized computational environments. To the best of our knowledge, this is the first distributed method using a single communication round in each iteration to jointly optimize the oracle and communication complexities for the problem considered here.
As a classical NP-hard problem and the topic of the PACE 2022 competition, the directed feedback vertex set problem (DFVSP) aims to find a minimum subset of vertices such that, when vertices in the subset and all their adjacent edges are removed from the directed graph, the remainder graph is acyclic. In this paper, we propose a threshold-based responsive simulated annealing algorithm called TRSA for solving DFVSP. First, we simplify the problem instances with two new reduction rules proposed in this paper and eight reduction rules from the literature. Then, based on a new solution representation, TRSA solves DFVSP with a fast local search procedure featured by a swap-based neighborhood structure and three neighborhood acceleration strategies. Finally, all these strategies are incorporated into a threshold-based responsive simulated annealing framework. Computational experiments on 140 benchmark instances show that TRSA is highly competitive compared to the state-of-the-art methods. Specifically, TRSA can improve the best known results for 53 instances, while matching the best known results for 79 ones. Furthermore, some important features of TRSA are analyzed to identify its success factors.