Research

Many large-scale machine learning and data mining tasks reduce to maximizing a submodular function under constraints—but the data is too large to fit in memory or must be processed in a distributed setting. We develop streaming and parallel algorithms that achieve near-optimal approximation guarantees with limited memory and few communication rounds.

Our contributions span monotone and non-monotone objectives under cardinality, matroid, and knapsack constraints. Key results include optimal single-pass streaming algorithms, the first MapReduce-compatible algorithms for non-monotone objectives, and adaptive complexity results that characterize the inherent parallelism of submodular optimization.

Key papers: Kuhnle, NeurIPS 2021; Kuhnle, NeurIPS 2019; Kuhnle et al., ICML 2018; Kuhnle, JAIR 2024; Kuhnle et al., KDD 2018; Kuhnle, AAAI 2019
See full publication list for complete coverage.

Modern combinatorial optimization instances can have billion-scale ground sets, yet optimal solutions use only a handful of elements. We develop pruning algorithms that quickly reduce the ground set to a manageable subset while provably preserving near-optimal solutions—a property we call containment. The pruned set can then be passed to any downstream optimizer, including exact solvers that would be intractable on the original instance.

Our work establishes the first theoretical framework for pruning complexity: information-theoretic limits on achievable containment ratios, algorithms that match these limits, and practical implementations. Key results include the first pruning algorithms for non-monotone submodular objectives and hardness results showing that extra pruning budget cannot improve containment beyond the greedy baseline for monotone functions.

Key papers: Nath & Kuhnle, AISTATS 2025; Kuhnle, AISTATS 2021 (Oral); Nath & Kuhnle, SoCS 2025

We develop machine learning pipelines for detecting and quantifying recreational fishing activity from satellite, AIS, and mobility data. This work, in collaboration with NOAA and LGL Ecological Research Associates, supports fisheries management by providing independent estimates of fishing effort across large spatial scales—data that is currently collected through costly creel surveys and phone interviews.

The pipeline integrates AIS vessel tracking, GPS trajectory analysis, and Cuebiq mobility data through object detection and activity classification models, producing spatiotemporal maps of recreational fishing pressure. The system is deployed for ongoing monitoring of Gulf of Mexico fisheries.

Collaborator: Nathan Putman (LGL Ecological Research Associates); funded by NOAA
In the news: Fishing for Better Tracking: Tracing Anglers in the Gulf Coast (TAMU Engineering News, April 2026) | Texas A&M University Researcher Tracks Recreational Fishing Using Machine Learning (Galveston Daily News)

When autonomous agents share a joint objective, how should each agent's contribution be valued, and how should the team coordinate? We study this through cooperative game theory—specifically, the Harsanyi dividend (Shapley value) decomposition of submodular reward functions. This provides a principled credit assignment: each agent receives credit proportional to its marginal contribution to every coalition it participates in.

Our coco-hs framework uses this decomposition to derive decentralized policies for cooperative multi-agent reinforcement learning, where agents optimize submodular objectives without centralized coordination.

Key paper: Chen, Richley, Perez-Lavin, Singh Syal, Kia & Kuhnle, JAAMAS (major revision)