Many statistical analyses, in both observational data and randomized control trials, ask: how does the outcome of interest vary with combinations of observable covariates? How do various drug combinations affect health outcomes, or how does technology adoption depend on incentives and demographics? Our goal is to partition this factorial space into ``pools'' of covariate combinations where the outcome differs across the pools (but not within a pool). Existing approaches (i) search for a single ``optimal'' partition under assumptions about the association between covariates or (ii) sample from the entire set of possible partitions. Both these approaches ignore the reality that, especially with correlation structure in covariates, many ways to partition the covariate space may be statistically indistinguishable, despite very different implications for policy or science. We develop an alternative perspective, called Rashomon Partition Sets (RPSs). Each item in the RPS partitions the space of covariates using a tree-like geometry. RPSs incorporate all partitions that have posterior values near the maximum a posteriori partition, even if they offer substantively different explanations, and do so using a prior that makes no assumptions about associations between covariates. This prior is the $\ell_0$ prior, which we show is minimax optimal. Given the RPS we calculate the posterior of any measurable function of the feature effects vector on outcomes, conditional on being in the RPS. We also characterize approximation error relative to the entire posterior and provide bounds on the size of the RPS. Simulations demonstrate this framework allows for robust conclusions relative to conventional regularization techniques. We apply our method to three empirical settings: price effects on charitable giving, chromosomal structure (telomere length), and the introduction of microfinance.

We study the problem of adversarial combinatorial bandit with a switching cost $\lambda$ for a switch of each selected arm in each round, considering both the bandit feedback and semi-bandit feedback settings. In the oblivious adversarial case with $K$ base arms and time horizon $T$, we derive lower bounds for the minimax regret and design algorithms to approach them. To prove these lower bounds, we design stochastic loss sequences for both feedback settings, building on an idea from previous work in Dekel et al. (2014). The lower bound for bandit feedback is $ \tilde{\Omega}\big( (\lambda K)^{\frac{1}{3}} (TI)^{\frac{2}{3}}\big)$ while that for semi-bandit feedback is $ \tilde{\Omega}\big( (\lambda K I)^{\frac{1}{3}} T^{\frac{2}{3}}\big)$ where $I$ is the number of base arms in the combinatorial arm played in each round. To approach these lower bounds, we design algorithms that operate in batches by dividing the time horizon into batches to restrict the number of switches between actions. For the bandit feedback setting, where only the total loss of the combinatorial arm is observed, we introduce the Batched-Exp2 algorithm which achieves a regret upper bound of $\tilde{O}\big((\lambda K)^{\frac{1}{3}}T^{\frac{2}{3}}I^{\frac{4}{3}}\big)$ as $T$ tends to infinity. In the semi-bandit feedback setting, where all losses for the combinatorial arm are observed, we propose the Batched-BROAD algorithm which achieves a regret upper bound of $\tilde{O}\big( (\lambda K)^{\frac{1}{3}} (TI)^{\frac{2}{3}}\big)$.

This paper presents an incremental replanning algorithm, dubbed LTL-D*, for temporal-logic-based task planning in a dynamically changing environment. Unexpected changes in the environment may lead to failures in satisfying a task specification in the form of a Linear Temporal Logic (LTL). In this study, the considered failures are categorized into two classes: (i) the desired LTL specification can be satisfied via replanning, and (ii) the desired LTL specification is infeasible to meet strictly and can only be satisfied in a "relaxed" fashion. To address these failures, the proposed algorithm finds an optimal replanning solution that minimally violates desired task specifications. In particular, our approach leverages the D* Lite algorithm and employs a distance metric within the synthesized automaton to quantify the degree of the task violation and then replan incrementally. This ensures plan optimality and reduces planning time, especially when frequent replanning is required. Our approach is implemented in a robot navigation simulation to demonstrate a significant improvement in the computational efficiency for replanning by two orders of magnitude.

Language models are rarely shown fruitful mistakes while training. They then struggle to look beyond the next token, suffering from a snowballing of errors and struggling to predict the consequence of their actions several steps ahead. In this paper, we show how language models can be taught to search by representing the process of search in language, as a flattened string -- a stream of search (SoS). We propose a unified language for search that captures an array of different symbolic search strategies. We demonstrate our approach using the simple yet difficult game of Countdown, where the goal is to combine input numbers with arithmetic operations to reach a target number. We pretrain a transformer-based language model from scratch on a dataset of streams of search generated by heuristic solvers. We find that SoS pretraining increases search accuracy by 25% over models trained to predict only the optimal search trajectory. We further finetune this model with two policy improvement methods: Advantage-Induced Policy Alignment (APA) and Self-Taught Reasoner (STaR). The finetuned SoS models solve 36% of previously unsolved problems, including problems that cannot be solved by any of the heuristic solvers. Our results indicate that language models can learn to solve problems via search, self-improve to flexibly use different search strategies, and potentially discover new ones.

Machine learning methods often assume that the test data have the same distribution as the training data. However, this assumption may not hold due to multiple levels of heterogeneity in applications, raising issues in algorithmic fairness and domain generalization. In this work, we address the problem of fair and generalizable machine learning by invariant principles. We propose a training environment-based oracle, FAIRM, which has desirable fairness and domain generalization properties under a diversity-type condition. We then provide an empirical FAIRM with finite-sample theoretical guarantees under weak distributional assumptions. We then develop efficient algorithms to realize FAIRM in linear models and demonstrate the nonasymptotic performance with minimax optimality. We evaluate our method in numerical experiments with synthetic data and MNIST data and show that it outperforms its counterparts.

Generative models are rapidly gaining popularity and being integrated into everyday applications, raising concerns over their safety issues as various vulnerabilities are exposed. Faced with the problem, the field of red teaming is experiencing fast-paced growth, which highlights the need for a comprehensive organization covering the entire pipeline and addressing emerging topics for the community. Our extensive survey, which examines over 120 papers, introduces a taxonomy of fine-grained attack strategies grounded in the inherent capabilities of language models. Additionally, we have developed the searcher framework that unifies various automatic red teaming approaches. Moreover, our survey covers novel areas including multimodal attacks and defenses, risks around multilingual models, overkill of harmless queries, and safety of downstream applications. We hope this survey can provide a systematic perspective on the field and unlock new areas of research. Warning: This paper contains examples that may be offensive, harmful, or biased.

Quadratic Assignment Problem (QAP) is a practical combinatorial optimization problems that has been studied for several years. Since it is NP-hard, solving large problem instances of QAP is challenging. Although heuristics can find semi-optimal solutions, the execution time significantly increases as the problem size increases. Recently, solving combinatorial optimization problems by deep learning has been attracting attention as a faster solver than heuristics. Even with deep learning, however, solving large QAP is still challenging. In this paper, we propose the deep reinforcement learning model called the two-stage graph pointer network (GPN) for solving QAP. Two-stage GPN relies on GPN, which has been proposed for Euclidean Traveling Salesman Problem (TSP). First, we extend GPN for general TSP, and then we add new algorithms to that model for solving QAP. Our experimental results show that our two-stage GPN provides semi-optimal solutions for benchmark problem instances from TSPlib and QAPLIB.

Neural architecture search (NAS) is an effective method for discovering new convolutional neural network (CNN) architectures. However, existing approaches often require time-consuming training or intensive sampling and evaluations. Zero-shot NAS aims to create training-free proxies for architecture performance prediction. However, existing proxies have suboptimal performance, and are often outperformed by simple metrics such as model parameter counts or the number of floating-point operations. Besides, existing model-based proxies cannot be generalized to new search spaces with unseen new types of operators without golden accuracy truth. A universally optimal proxy remains elusive. We introduce TG-NAS, a novel model-based universal proxy that leverages a transformer-based operator embedding generator and a graph convolution network (GCN) to predict architecture performance. This approach guides neural architecture search across any given search space without the need of retraining. Distinct from other model-based predictor subroutines, TG-NAS itself acts as a zero-cost (ZC) proxy, guiding architecture search with advantages in terms of data independence, cost-effectiveness, and consistency across diverse search spaces. Our experiments showcase its advantages over existing proxies across various NAS benchmarks, suggesting its potential as a foundational element for efficient architecture search. TG-NAS achieves up to 300X improvements in search efficiency compared to previous SOTA ZC proxy methods. Notably, it discovers competitive models with 93.75% CIFAR-10 accuracy on the NAS-Bench-201 space and 74.5% ImageNet top-1 accuracy on the DARTS space.

Multi-agent path finding (MAPF) is the problem of finding collision-free paths for a team of agents to reach their goal locations. State-of-the-art classical MAPF solvers typically employ heuristic search to find solutions for hundreds of agents but are typically centralized and can struggle to scale when run with short timeouts. Machine learning (ML) approaches that learn policies for each agent are appealing as these could enable decentralized systems and scale well while maintaining good solution quality. Current ML approaches to MAPF have proposed methods that have started to scratch the surface of this potential. However, state-of-the-art ML approaches produce "local" policies that only plan for a single timestep and have poor success rates and scalability. Our main idea is that we can improve a ML local policy by using heuristic search methods on the output probability distribution to resolve deadlocks and enable full horizon planning. We show several model-agnostic ways to use heuristic search with learnt policies that significantly improve the policies' success rates and scalability. To our best knowledge, we demonstrate the first time ML-based MAPF approaches have scaled to high congestion scenarios (e.g. 20% agent density).

We study the generalization capability of Unsupervised Learning in solving the Travelling Salesman Problem (TSP). We use a Graph Neural Network (GNN) trained with a surrogate loss function to generate an embedding for each node. We use these embeddings to construct a heat map that indicates the likelihood of each edge being part of the optimal route. We then apply local search to generate our final predictions. Our investigation explores how different training instance sizes, embedding dimensions, and distributions influence the outcomes of Unsupervised Learning methods. Our results show that training with larger instance sizes and increasing embedding dimensions can build a more effective representation, enhancing the model's ability to solve TSP. Furthermore, in evaluating generalization across different distributions, we first determine the hardness of various distributions and explore how different hardnesses affect the final results. Our findings suggest that models trained on harder instances exhibit better generalization capabilities, highlighting the importance of selecting appropriate training instances in solving TSP using Unsupervised Learning.