Conventional methods for query autocompletion aim to predict which completed query a user will select from a list. A shortcoming of this approach is that users often do not know which query will provide the best retrieval performance on the current information retrieval system, meaning that any query autocompletion methods trained to mimic user behavior can lead to suboptimal query suggestions. To overcome this limitation, we propose a new approach that explicitly optimizes the query suggestions for downstream retrieval performance. We formulate this as a problem of ranking a set of rankings, where each query suggestion is represented by the downstream item ranking it produces. We then present a learning method that ranks query suggestions by the quality of their item rankings. The algorithm is based on a counterfactual learning approach that is able to leverage feedback on the items (e.g., clicks, purchases) to evaluate query suggestions through an unbiased estimator, thus avoiding the assumption that users write or select optimal queries. We establish theoretical support for the proposed approach and provide learning-theoretic guarantees. We also present empirical results on publicly available datasets, and demonstrate real-world applicability using data from an online shopping store.

Empirical results in software engineering have long started to show that findings are unlikely to be applicable to all software systems, or any domain: results need to be evaluated in specified contexts, and limited to the type of systems that they were extracted from. This is a known issue, and requires the establishment of a classification of software types. This paper makes two contributions: the first is to evaluate the quality of the current software classifications landscape. The second is to perform a case study showing how to create a classification of software types using a curated set of software systems. Our contributions show that existing, and very likely even new, classification attempts are deemed to fail for one or more issues, that we named as the `antipatterns' of software classification tasks. We collected 7 of these antipatterns that emerge from both our case study, and the existing classifications. These antipatterns represent recurring issues in a classification, so we discuss practical ways to help researchers avoid these pitfalls. It becomes clear that classification attempts must also face the daunting task of formulating a taxonomy of software types, with the objective of establishing a hierarchy of categories in a classification.

Publicly traded companies are required to submit periodic reports with eXtensive Business Reporting Language (XBRL) word-level tags. Manually tagging the reports is tedious and costly. We, therefore, introduce XBRL tagging as a new entity extraction task for the financial domain and release FiNER-139, a dataset of 1.1M sentences with gold XBRL tags. Unlike typical entity extraction datasets, FiNER-139 uses a much larger label set of 139 entity types. Most annotated tokens are numeric, with the correct tag per token depending mostly on context, rather than the token itself. We show that subword fragmentation of numeric expressions harms BERT's performance, allowing word-level BILSTMs to perform better. To improve BERT's performance, we propose two simple and effective solutions that replace numeric expressions with pseudo-tokens reflecting original token shapes and numeric magnitudes. We also experiment with FIN-BERT, an existing BERT model for the financial domain, and release our own BERT (SEC-BERT), pre-trained on financial filings, which performs best. Through data and error analysis, we finally identify possible limitations to inspire future work on XBRL tagging.

He has been involved in many areas that make use of AI and ML at EA, particularly AI for games development and verification. He started out with game development, but is now with the AI support team, which supports all of the company's teams.

Most convergence guarantees for stochastic gradient descent with momentum (SGDm) rely on iid sampling. Yet, SGDm is often used outside this regime, in settings with temporally correlated input samples such as continual learning and reinforcement learning. Existing work has shown that SGDm with a decaying step-size can converge under Markovian temporal correlation. In this work, we show that SGDm under covariate shift with a fixed step-size can be unstable and diverge. In particular, we show SGDm under covariate shift is a parametric oscillator, and so can suffer from a phenomenon known as resonance. We approximate the learning system as a time varying system of ordinary differential equations, and leverage existing theory to characterize the system's divergence/convergence as resonant/nonresonant modes. The theoretical result is limited to the linear setting with periodic covariate shift, so we empirically supplement this result to show that resonance phenomena persist even under non-periodic covariate shift, nonlinear dynamics with neural networks, and optimizers other than SGDm.

In the multi-agent path finding problem (MAPF) we are given a set of agents each with respective start and goal positions. The task is to find paths for all agents while avoiding collisions, aiming to minimize a given objective function. Many MAPF solvers were introduced in the past decade for optimizing two specific objective functions: sum-of-costs and makespan. Two prominent categories of solvers can be distinguished: search-based solvers and compilation-based solvers. Search-based solvers were developed and tested for the sum-of-costs objective, while the most prominent compilation-based solvers that are built around Boolean satisfiability (SAT) were designed for the makespan objective. Very little is known on the performance and relevance of solvers from the compilation-based approach on the sum-of-costs objective. In this paper, we start to close the gap between these cost functions in the compilation-based approach. Our main contribution is a new SAT-based MAPF solver called MDD-SAT, that is directly aimed to optimally solve the MAPF problem under the sum-of-costs objective function. Using both a lower bound on the sum-of-costs and an upper bound on the makespan, MDD-SAT is able to generate a reasonable number of Boolean variables in our SAT encoding. We then further improve the encoding by borrowing ideas from ICTS, a search-based solver. In addition, we show that concepts applicable in search-based solvers like ICTS and ICBS are applicable in the SAT-based approach as well. Specifically, we integrate independence detection, a generic technique for decomposing an MAPF instance into independent subproblems, into our SAT-based approach, and we design a relaxation of our optimal SAT-based solver that results in a bounded suboptimal SAT-based solver. Experimental evaluation on several domains shows that there are many scenarios where our SAT-based methods outperform state-of-the-art sum-of-costs search-based solvers, such as variants of the ICTS and ICBS algorithms.

To achieve the ambitious goals of artificial intelligence, reinforcement learning must include planning with a model of the world that is abstract in state and time. Deep learning has made progress in state abstraction, but, although the theory of time abstraction has been extensively developed based on the options framework, in practice options have rarely been used in planning. One reason for this is that the space of possible options is immense and the methods previously proposed for option discovery do not take into account how the option models will be used in planning. Options are typically discovered by posing subsidiary tasks such as reaching a bottleneck state, or maximizing a sensory signal other than the reward. Each subtask is solved to produce an option, and then a model of the option is learned and made available to the planning process. The subtasks proposed in most previous work ignore the reward on the original problem, whereas we propose subtasks that use the original reward plus a bonus based on a feature of the state at the time the option stops. We show that options and option models obtained from such reward-respecting subtasks are much more likely to be useful in planning and can be learned online and off-policy using existing learning algorithms. Reward respecting subtasks strongly constrain the space of options and thereby also provide a partial solution to the problem of option discovery. Finally, we show how the algorithms for learning values, policies, options, and models can be unified using general value functions.

The policy gradient theorem (Sutton et al., 2000) prescribes the usage of a cumulative discounted state distribution under the target policy to approximate the gradient. Most algorithms based on this theorem, in practice, break this assumption, introducing a distribution shift that can cause the convergence to poor solutions. In this paper, we propose a new approach of reconstructing the policy gradient from the start state without requiring a particular sampling strategy. The policy gradient calculation in this form can be simplified in terms of a gradient critic, which can be recursively estimated due to a new Bellman equation of gradients. By using temporal-difference updates of the gradient critic from an off-policy data stream, we develop the first estimator that sidesteps the distribution shift issue in a model-free way. We prove that, under certain realizability conditions, our estimator is unbiased regardless of the sampling strategy. We empirically show that our technique achieves a superior bias-variance trade-off and performance in presence of off-policy samples.

Time series forecasting is fundamental for various use cases in different domains such as energy systems and economics. Creating a forecasting model for a specific use case requires an iterative and complex design process. The typical design process includes the five sections (1) data pre-processing, (2) feature engineering, (3) hyperparameter optimization, (4) forecasting method selection, and (5) forecast ensembling, which are commonly organized in a pipeline structure. One promising approach to handle the ever-growing demand for time series forecasts is automating this design process. The present paper, thus, analyzes the existing literature on automated time series forecasting pipelines to investigate how to automate the design process of forecasting models. Thereby, we consider both Automated Machine Learning (AutoML) and automated statistical forecasting methods in a single forecasting pipeline. For this purpose, we firstly present and compare the proposed automation methods for each pipeline section. Secondly, we analyze the automation methods regarding their interaction, combination, and coverage of the five pipeline sections. For both, we discuss the literature, identify problems, give recommendations, and suggest future research. This review reveals that the majority of papers only cover two or three of the five pipeline sections. We conclude that future research has to holistically consider the automation of the forecasting pipeline to enable the large-scale application of time series forecasting.

Testing with quiz questions has proven to be an effective way to assess and improve the educational process. However, manually creating quizzes is tedious and time-consuming. To address this challenge, we present Leaf, a system for generating multiple-choice questions from factual text. In addition to being very well suited for the classroom, Leaf could also be used in an industrial setting, e.g., to facilitate onboarding and knowledge sharing, or as a component of chatbots, question answering systems, or Massive Open Online Courses (MOOCs). The code and the demo are available on GitHub.