Publications & Projects
Patent
Systems and Methods for Associating Data With a Non-Material Concept
Systems and methods for associating data with non-material concepts that allows for subsequent searching and retrieval of such non-material concepts. A concept definition is received that defines a concept identifier and a plurality of discrete stages. Retrieved data is analyzed to determine correspondence with the concept and its stages, enabling holistic concept-based search and retrieval.
Doctoral Thesis
Objective Measures of Complexity for Dynamic Decision-Making Problems
Introduced a quantitative framework for characterizing and predicting human decision-making in dynamic, uncertain environments. Formalized ten objective measures of decision complexity based on computational, system, and cognitive complexity theories. Combined simulation-based experimentation with data-analytic approaches to model learning, adaptation, and metacognitive control—anticipating current trends in reinforcement learning, probabilistic modeling, and human-AI teaming.
Refereed Conference Papers & Journal Articles
Constant, A., Westermann, H., Wilson, B., Kiefer, A., Hipolito, I., Pronovost, S., Swanson, S., Albarracin, M., Ramstead, M. J.D. A Path Towards Legal Autonomy: An interoperable and explainable approach to extracting, transforming, loading and computing legal information using large language models, expert systems and Bayesian networks. Journal of Artificial Intelligence and Law.
Chatterjee, B., Pronovost, S., Brassard, H., & Patel, B. Creating geospecific synthetic environments using deep learning and process automation. I/ITSEC Proceedings.
Lafond, D., Gagnon, J. F., Pronovost, S., Ducharme, M., & Tremblay, S. Behavioral Test for Prediction of Individual Differences in Dynamic Decision Making Ability. 7th International Conference on Applied Human Factors and Ergonomics.
Lafond, D., Gagnon, J.-F., St-Louis, M.-È., Pronovost, S., DuCharme, M. B., & Tremblay, S. Decision Heuristics and Human Performance in a Policy Management Simulation. 32nd International Conference of the System Dynamics Society.
Pronovost, S., Gagnon, J.-F., Lafond, D., & Tremblay, S. Objective measures of complexity for dynamic decision making in an interactive learning environment. 1st Asia-Pacific System Dynamics Conference.
Pronovost, S. Using Micro-Architectures of Cognition to Model Macro-Cognitive Systems: A Warfighting Video Game Example. 21st BRIMS Conference.
Murray, J. L., Hagen, L., Pronovost, S., & Lai, G. Ottawa Paramedic Service Communication Centre: Workload Analysis and Predictive Modeling. 4th International Conference on Applied Human Factors and Ergonomics.
Pronovost, S. Improving Usability and Integration of Human Behaviour Representation Engineering across Cognitive Modeling, Human Factors, and Modeling and Simulation Best Practices. 19th BRIMS Conference.
West, R. L., & Pronovost, S. Modeling SGOMS in ACT-R: Linking Macro- and Micro-cognition. Journal of Cognitive Engineering and Decision Making, 3(2), 194–207.
Pronovost, S. Bridging Cognitive Modeling and Human Behaviour Representation Engineering. 9th Conference of the Australasian Society for Cognitive Science.
Crebolder, J. M., & Pronovost, S. Investigating Virtual Social Networking in the Military Domain. 14th ICCRTS.
Pronovost, S., & West, R. L. Bridging cognitive modeling and model-based evaluation: extending GOMS to model virtual sociotechnical systems and strategic activities. 52nd Annual Meeting of HFES.
Selected Academic Poster Presentations
Pronovost, S., St-Louis, M.-E., Lafond, D., Gagnon, J.-F., DuCharme, M. B., & Tremblay, S. Implicit learning in dynamic decision making: A glass-box approach. CogSci 2015, Los Angeles, CA.
Pronovost, S., Gagnon, J.-F., Lafond, D., & Tremblay, S. Models of Complexity for Dynamic Decision Making Derived from Cognitive Informatics, Complexity Theory, and Psychophysics. CogSci 2014, Québec City.
Pronovost, S., & West, R. L. A GOMS Model of Virtual Sociotechnical Systems: Using Video Games to Build Cognitive Models. European Conference on Cognitive Ergonomics.
Selected Technical Reports
Over 30 technical reports authored or co-authored for defence, government, and industry stakeholders including DRDC Canada, Lockheed Martin, CAE, and Thales. Selected highlights:
- Lafond, D., Pronovost, S., et al. (2017). Improving complex problem solving through collaborative learning and ITS support.
- Banbury, S., Gagnon, J.-F., & Pronovost, S. (2017). HF Support to Evaluation of GCS Workspace Options. For DRDC Toronto.
- Pronovost, S. & Banbury, S. (2016). Performance and Workload Modeling Report for the Pattern of Life Authority Pathway.
- Banbury, S., Pronovost, S., et al. (2015). UAS Operator Task Analysis Refinement and Performance Modeling.
- Tremblay, S., Vachon, F., Gagnon, J.-F., Pronovost, S., & Banbury, S. (2013). Cognitive Training for Emergency Response: SYnRGY as an Intelligent Serious Game.
- Zobarich, R., Pronovost, S., et al. (2011). Dynamic Synthetic Environment & Human Behaviour Representation Analysis.
- Banbury, S., Baker, K., et al. (2009). Halifax Class Modernization Program: Human Factors Review. For Lockheed Martin / DND.
- Pronovost, S. & Lai, G. (2009). Visual Perception Modeling for Virtual Operators. For DRDC Toronto.
Project Highlights
Probabilistic Programming for Active Inference Systems
Led the design and implementation of a high-performance probabilistic programming layer grounded in constrained Forney-style factor graphs (CFFGs). The system introduced a variational message passing (VMP) engine optimized in JAX, leveraging a fully vectorized, single flat-array representation for belief propagation and parameter learning across entire model graphs.
Results
Delivered a JAX-native, differentiable probabilistic inference backend scaling to thousands of factors and variables in real time, with significant efficiency gains through array-level vectorization and JIT compilation.
Significance
Established a novel bridge between probabilistic programming, active inference, and computational neuroscience, laying the foundation for hybrid symbolic–neural architectures for hierarchical and distributed intelligence systems.
Deep Learning Pipeline Automation for Warranty Analytics
Led the design and deployment of a unified ML engineering platform combining Kedro for pipeline orchestration and MLflow for experiment tracking. Built and fine-tuned PyTorch & Hugging Face–based LLMs for document classification, semantic search, and warranty-claims triage.
Results
Reduced model-deployment time by over 20% and improved reproducibility across business units. The Kedro–MLflow framework became the organization's reference for standardized, auditable ML workflows.
Significance
Unified LLM-based document intelligence, generative AI, and MLOps automation within a single reproducible architecture.
Deep Learning Applied to Computer Vision with Synthetic Data
Developed a hybrid and synthetic data creation pipeline conflating multi-spectral satellite/aerial imagery with procedural content creation for deep learning computer vision tasks. Trained ICT-Net (U-Net architecture with dense residual and squeeze-and-excitation blocks) on sensor, synthetic, and hybrid datasets.
Results
Demonstrated how aligning data synthesis with real-world tasks can mitigate costly data-acquisition barriers, with future phases aimed at ensembling multispectral DNNs for autonomous systems guidance and control.
Significance
Capitalized on the best features of M&S technology and autonomous AI systems, using simulation-ready synthetic environments to train AI models and computer vision AI to automate large-scale synthetic environments reconstruction.
ML and Deep Learning Applied to Natural Language Processing
Developed a ML capability for semantic search in biotechnology, pharmaceutics, and life sciences research. Combined unsupervised learning (word2vec, doc2vec, Gensim), self-supervised learning (autoencoders), and supervised learning (deep RNNs) to produce candidates with latent semantic similarity for knowledge management.
Results
Developed an iterative tf-idf scheme to improve signal-to-noise ratio in latent semantic analysis, combined with latent Dirichlet allocation for topic modeling and multi-label classification via recurrent neural networks.
Significance
Demonstrated how deep-learning-powered latent semantic analysis offers considerable improvements for information retrieval in knowledge management systems.
Machine Learning Applied to Quantitative Data Analysis
Deployed machine learning approaches (Python, R, scikit-learn, caret) for unsupervised and supervised learning models on behavioral performance data from dynamic decision-making experiments. The noisy and sparse datasets were unsuitable for traditional statistical hypothesis testing.
Results
Advanced multivariate methods including non-parametric modeling (MM-estimation using IRLS) performed well for predictive models compared to basic linear approaches.
Significance
A rare application of machine learning approaches to data analysis in behavioral sciences, providing external validation of ML tools for behavioral data analysis.