Stephen Josef Wissow

Photo of Stephen Wissow

I am a doctoral student in the Department of Computer Science at the University of New Hampshire where I work with Professor Wheeler Ruml in the UNH Artificial Intelligence Group, and I also collaborate with Dr. Masataro Asai from the MIT-IBM Watson AI Lab in Cambridge, MA. My research interests are in algorithms for planning and autonomy. I completed my MS degree in computer science in 2020 with research in the area of network time synchronization and one-way latency measurement with Professor Radim Bartoš.

See my papers and current research below (updated March 2024). You can find my CV here (updated March 2024).

Please feel free to reach out! sjw[at]cs[dot]unh[dot]edu / LinkedIn / GitHub

Papers

Stephen Wissow, Fanhao Yu, and Wheeler Ruml, Tunable Suboptimal Heuristic Search, Proceedings of the Seventeenth International Symposium on Combinatorial Search (SoCS-24), 2024.

Stephen Wissow and Masataro Asai, Scale-Adaptive Balancing of Exploration and Exploitation in Classical Planning, Proceedings of the ICAPS-23 Workshop on Heuristics and Search for Domain-Independent Planning (HSDIP-23), 2023. slides

Current Research

Tunable Suboptimal Heuristic Search (Fall 2023--present, Prof. Wheeler Ruml): achieving the best empirical cost-vs-time trade-off when tuning suboptimality to achieve desired running time. Comparative analysis of bounded and unbounded suboptimal algorithms, and presentation of a new, robust tunable algorithm that appears to fare better than all others on domains with dead-end states.

Distributional Heuristics: Balancing Exploration and Exploitation in Agile Search (Summer 2022--present, Dr. Masataro Asai): In joint collaboration with Dr. Asai, this work presents a new Multi-Armed Bandit (MAB), UCB1-Normal2, for agile planning (satisficing planning that minimizes planning time with zero regard for solution cost) within a Monte Carlo Tree Search (MCTS) framework that outperforms previous related work and the traditional baseline of Greedy Best-First Search (GBFS) on a range of International Planning Competition (IPC) benchmarks. Future work will evaluate in terms of CPU time instead of expansions, and consider our new distributional heuristic, hFFrand, which exhibits variation in heuristic value on a majority of domains tested (here showing initial state heuristic value for increasingly difficult instances per domain):
Planning with a Known Deadline (Fall 2021--present, Prof. Wheeler Ruml): Investigating online prediction of future search effort in order to take greatest advantage of planning time remaining prior to a known deadline.

Previous Research

Motion Planning with Dynamic Obstacles (Fall 2020--Fall 2021, Prof. Wheeler Ruml): Implementing offline anytime BIT* and integrating into a field-tested ROS-based autonomy stack for a surface vessel for performance comparison with alternative realtime algorithm.

BIT* motion tree with solution avoiding static and dynamic obstacles

Screenshot of ASV avoiding static and dynimac obstacles by using BIT* in CAMP

Time Enough: Synchronization for Latency Measurement, MS thesis, University of New Hampshire, 2020 (advisor: Radim Bartoš): Empirical investigation of relative robustness of several network time synchronization methods with respect to asymmetric latency and asymmetric jitter for the purpose of one-way latency measurement.

Diagram of asymmetric latency in communication

Boxplots showing Precise Time Protocol (PTP) with least relative difference from GPS ground truth in one-way latency when subject to asymmetric jitter

Class Projects

Planning with a Known Deadline (CS 931: Planning for Robots, Fall 2021, Prof. Wheeler Ruml): Comparing different heuristics and error models for use in a deadline-aware version of beam search.
Data-Driven Dynamics Model for a Boat (CS 850: Machine Learning, Spring 2021, Prof. Marek Petrik): Used k-nearest neighbors to learn a data-driven model from physics-based simulated data as proof-of-concept for use in model-predictive control.