Searcharxiv⌕ Search

arXiv subjects

Thomas Robinson

Publications and source records attributed to Thomas Robinson.

10 recordsLinked to original sources
arXiv2026

Fault-Tolerant Design and Multi-Objective Model Checking for Real-Time Deep Reinforcement Learning Systems

Deep reinforcement learning (DRL) has emerged as a powerful paradigm for solving complex decision-making problems. However, DRL-based systems still face significant dependability challenges particularly in real-time environments due to the simulation-to-reality gap, out-of-distribution observations, and the critical impact of latency. Latency-induced faults, in particular, can lead to unsafe or unstable behaviour, yet existing fault-tolerance approaches to DRL systems lack formal methods to rigorously analyse and optimise performance and safety simultaneously in real-time settings. To address this, we propose a formal framework for designing and analysing real-time switching mechanisms between DRL agents and alternative controllers. Our approach leverages Timed Automata (TAs) for explicit switch logic design, which is then syntactically converted to a Markov Decision Process (MDP) for formal analysis. We develop a novel convex query technique for multi-objective model checking, enabling the optimisation of soft performance objectives while ensuring hard safety constraints for MDPs. Furthermore, we present MOPMC, a GPU-accelerated software tool implementing this technique, demonstrating superior scalability in both model size and objective numbers.

cs.SE↗
arXiv2026

Learning Next Action Predictors from Human-Computer Interaction

Truly proactive AI systems must anticipate what we will do next. This foresight demands far richer information than the sparse signals we type into our prompts -- it demands reasoning over the entire context of what we see and do. We formalize this as next action prediction (NAP): given a sequence of a user's multimodal interactions with a computer (screenshots, clicks, sensor data), predict that user's next action. Progress on this task requires both new data and modeling approaches. To scale data, we annotate longitudinal, naturalistic computer use with vision-language models. We release an open-source pipeline for performing this labeling on private infrastructure, and label over 360K actions across one month of continuous phone usage from 20 users, amounting to 1,800 hours of screen time. We then introduce LongNAP, a user model that combines parametric and in-context learning to reason over long interaction histories. LongNAP is trained via policy gradient methods to generate user-specific reasoning traces given some context; retrieve relevant traces from a library of past traces; and then apply retrieved traces in-context to predict future actions. Using an LLM-as-judge evaluation metric (0-1 similarity to ground truth), LongNAP significantly outperforms supervised finetuning and prompted baselines on held-out data (by 79% and 39% respectively). Additionally, LongNAP generalizes to held out users when trained across individuals. The space of next actions a user might take at any moment is unbounded, spanning thousands of possible outcomes. Despite this, 17.1% of LongNAP's predicted trajectories are well-aligned with what a user does next (LLM-judge score $\geq$ 0.5). This rises to 26% when we filter to highly confident predictions. In sum, we argue that learning from the full context of user behavior to anticipate user needs is now a viable task with substantial opportunity.

cs.CL↗
arXiv2023

Active learning with biased non-response to label requests

Active learning can improve the efficiency of training prediction models by identifying the most informative new labels to acquire. However, non-response to label requests can impact active learning's effectiveness in real-world contexts. We conceptualise this degradation by considering the type of non-response present in the data, demonstrating that biased non-response is particularly detrimental to model performance. We argue that biased non-response is likely in contexts where the labelling process, by nature, relies on user interactions. To mitigate the impact of biased non-response, we propose a cost-based correction to the sampling strategy--the Upper Confidence Bound of the Expected Utility (UCB-EU)--that can, plausibly, be applied to any active learning algorithm. Through experiments, we demonstrate that our method successfully reduces the harm from labelling non-response in many settings. However, we also characterise settings where the non-response bias in the annotations remains detrimental under UCB-EU for specific sampling methods and data generating processes. Finally, we evaluate our method on a real-world dataset from an e-commerce platform. We show that UCB-EU yields substantial performance improvements to conversion models that are trained on clicked impressions. Most generally, this research serves to both better conceptualise the interplay between types of non-response and model improvements via active learning, and to provide a practical, easy-to-implement correction that mitigates model degradation.

cs.LG↗
arXiv2023

Identification, Impacts, and Opportunities of Three Common Measurement Considerations when using Digital Trace Data

Cataloguing specific URLs, posts, and applications with digital traces is the new best practice for measuring media use and content consumption. Despite the apparent accuracy that comes with greater granularity, however, digital traces may introduce additional ambiguity and new errors into the measurement of media use. In this note, we identify three new measurement challenges when using Digital Trace Data that were recently uncovered using a new measurement framework - Screenomics - that records media use at the granularity of individual screenshots obtained every few seconds as people interact with mobile devices. We label the considerations as follows: (1) entangling - the common measurement error introduced by proxying exposure to content by exposure to format; (2) flattening - aggregating unique segments of media interaction without incorporating temporal information, most commonly intraindividually and (3) bundling - summation of the durations of segments of media interaction, indiscriminate with respect to variations across media segments.

cs.HC↗
arXiv2023

Multi-Objective Task Assignment and Multiagent Planning with Hybrid GPU-CPU Acceleration

Allocation and planning with a collection of tasks and a group of agents is an important problem in multiagent systems. One commonly faced bottleneck is scalability, as in general the multiagent model increases exponentially in size with the number of agents. We consider the combination of random task assignment and multiagent planning under multiple-objective constraints, and show that this problem can be decentralised to individual agent-task models. We present an algorithm of point-oriented Pareto computation, which checks whether a point corresponding to given cost and probability thresholds for our formal problem is feasible or not. If the given point is infeasible, our algorithm finds a Pareto-optimal point which is closest to the given point. We provide the first multi-objective model checking framework that simultaneously uses GPU and multi-core acceleration. Our framework manages CPU and GPU devices as a load balancing problem for parallel computation. Our experiments demonstrate that parallelisation achieves significant run time speed-up over sequential computation.

cs.MA↗
arXiv2020

On Biased Random Walks, Corrupted Intervals, and Learning Under Adversarial Design

We tackle some fundamental problems in probability theory on corrupted random processes on the integer line. We analyze when a biased random walk is expected to reach its bottommost point and when intervals of integer points can be detected under a natural model of noise. We apply these results to problems in learning thresholds and intervals under a new model for learning under adversarial design.

cs.LG↗
arXiv2019

The Norms of Graph Spanners

A $t$-spanner of a graph $G$ is a subgraph $H$ in which all distances are preserved up to a multiplicative $t$ factor. A classical result of Althöfer et al. is that for every integer $k$ and every graph $G$, there is a $(2k-1)$-spanner of $G$ with at most $O(n^{1+1/k})$ edges. But for some settings the more interesting notion is not the number of edges, but the degrees of the nodes. This spurred interest in and study of spanners with small maximum degree. However, this is not necessarily a robust enough objective: we would like spanners that not only have small maximum degree, but also have "few" nodes of "large" degree. To interpolate between these two extremes, in this paper we initiate the study of graph spanners with respect to the $\ell_p$-norm of their degree vector, thus simultaneously modeling the number of edges (the $\ell_1$-norm) and the maximum degree (the $\ell_{\infty}$-norm). We give precise upper bounds for all ranges of $p$ and stretch $t$: we prove that the greedy $(2k-1)$-spanner has $\ell_p$ norm of at most $\max(O(n), O(n^{(k+p)/(kp)}))$, and that this bound is tight (assuming the Erdős girth conjecture). We also study universal lower bounds, allowing us to give "generic" guarantees on the approximation ratio of the greedy algorithm which generalize and interpolate between the known approximations for the $\ell_1$ and $\ell_{\infty}$ norm. Finally, we show that at least in some situations, the $\ell_p$ norm behaves fundamentally differently from $\ell_1$ or $\ell_{\infty}$: there are regimes ($p=2$ and stretch $3$ in particular) where the greedy spanner has a provably superior approximation to the generic guarantee.

cs.DS↗
arXiv2019

Guess What's on my Screen? Clustering Smartphone Screenshots with Active Learning

A significant proportion of individuals' daily activities is experienced through digital devices. Smartphones in particular have become one of the preferred interfaces for content consumption and social interaction. Identifying the content embedded in frequently-captured smartphone screenshots is thus a crucial prerequisite to studies of media behavior and health intervention planning that analyze activity interplay and content switching over time. Screenshot images can depict heterogeneous contents and applications, making the a priori definition of adequate taxonomies a cumbersome task, even for humans. Privacy protection of the sensitive data captured on screens means the costs associated with manual annotation are large, as the effort cannot be crowd-sourced. Thus, there is need to examine utility of unsupervised and semi-supervised methods for digital screenshot classification. This work introduces the implications of applying clustering on large screenshot sets when only a limited amount of labels is available. In this paper we develop a framework for combining K-Means clustering with Active Learning for efficient leveraging of labeled and unlabeled samples, with the goal of discovering latent classes and describing a large collection of screenshot data. We tested whether SVM-embedded or XGBoost-embedded solutions for class probability propagation provide for more well-formed cluster configurations. Visual and textual vector representations of the screenshot images are derived and combined to assess the relative contribution of multi-modal features to the overall performance.

cs.CV↗
arXiv2004

Dreidel Lasts $O(N^2)$ Spins

Dreidel is a popular game played during the festival of Chanukah. Players start with an equal number of tokens, donate one token each to a common pot, and take turns spinning a four-sided top, called the dreidel. Depending on the side showing up, the spinner takes all the tokens in the pot, takes half the tokens in the pot, gives a token to the pot, or does nothing. Whenever the pot goes empty, everyone donates a token to the pot. The game continues till all players, except one, go broke. We prove that the expected length of a game of dreidel is at most quadratic in the number of tokens, irrespective of the number of players. This proves a conjecture of Doron Zeilberger.

math.CO↗