Back to search

University of Nevada - Reno

Deep Generative and Explainable Learning Frameworks for Intent Recognition in Naval Domain

Abstract

dc:description.abstract

Accurate recognition of vessel intent supports maritime situational awareness, autonomous navigation, and tactical decision support. Three obstacles limit progress: the scarcity of labeled hostile-encounter data, noisy and intermittent sensor streams, and the inability of data-driven models to encode symbolic navigation rules and expert heuristics. This dissertation addresses these obstacles through five contributions evaluated on a unified multi-vessel dataset. First, NavySim is a scriptable multi-agent maritime simulator that generates benign and adversarial encounters with per-timestep intent labels, an intent recognition pipeline integrated, COLREGs-compliant navigation planners, and threat-visualization tools, supplying labeled encounter data unavailable in public AIS corpora. Second, Categorical Permutation Feature Importance (CPFI) is a model-agnostic feature-attribution method that ranks features by importance under class imbalance and multi-level categorical predictors. Third, two intent-classification studies apply the CPFI/Transformer Feature Importance Score (TFIS, developed by a collaborator) pipeline: an intent classifier that detects vessel behavior from long observation windows, and a feature-optimized pipeline that uses attribution scores within the training loop to prune, weight, and re-rank features. Fourth, we evaluate classical Bayesian trajectory predictors, including the Unscented Kalman Filter (UKF), Extended Kalman Filter (EKF), and UKF with Constant Turn Rate and Velocity (CTRV) dynamics, are evaluated on the same dataset. This assessment aims to highlight the structural limitations of intent-agnostic filters in hostile scenarios. Additionally, it establishes the matched-configuration baselines for comparing learned and generative models. Fifth, MTITP (Multi-Task Intent and Trajectory Prediction) is designed as a unified architecture for joint scenario-intent classification, intent prediction, and intent-conditioned trajectory generation over a short observation window of vessel data. It comprises a three-stream encoder that fuses per-vessel kinematics, pairwise geometry and symbolic COLREGs-derived interaction features through alternating temporal and cross-vessel attention, intent-classification and intent-prediction heads, and a trajectory generator trained under a two-phase protocol. Three backbone variants, named MTITP-WGAN, MTITP-L2, and MarITGAN with an IntentAwareCritic, differ only in the adversarial trajectory loss applied during the joint phase. Across its three variants, MTITP outperforms the strongest classical Kalman trajectory baselines in displacement error and exceeds published deep maritime intent classifiers in past- and future-intent accuracy, with gains concentrated in hostile intent scenarios.

Degree

thesis:*
Level thesis:degree_level
Doctorate Degree
Department thesis:degree_department
Computer Science
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Sayed, Md Abu
Advisors dc:contributor.advisor
  • Nicolescu, Monica Prof.
  • Nicolescu, Mircea Prof.
Committee members dc:contributor.committeemember
  • Tavakkoli, Alireza Prof.
  • Dascalu, Sergiu Prof.
  • Ramazan, Birant Dr.

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en_US, English

Identifiers

dc:identifier.*
Repository record dc:identifier.uri
https://scholarwolf.unr.edu/handle/11714/11925

Chain of custody

source
Harvested from
University of Nevada - Reno
Base URL
scholarwolf.unr.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Sayed, Md Abu. Deep Generative and Explainable Learning Frameworks for Intent Recognition in Naval Domain. Doctorate Degree thesis, 2026. https://scholarwolf.unr.edu/handle/11714/11925