Kevin Osanlou

I am a Research Scientist and Consultant in Machine Learning at Talan.

My areas of expertise involve Machine Learning, Graph Machine Learning, Planning. I also have interests in Computer Vision and Natural Language Processing. 

I hold a PhD in Machine Learning and Automated Planning and Scheduling from Paris-Dauphine University.

I hold two master degrees in Computer Science:

• Paris-Dauphine,  Master in AI with Machine Learning major

• EISTI, Master in Data Science

Osanlou, Kevin, Jeremy Frank, J. Benton, Andrei Bursuc, Christophe Guettier, Eric Jacopin, and Tristan Cazenave . “Solving Disjunctive Temporal Networks with Uncertainty under Restricted Time-Based Controllability Using Tree Search and Graph Neural Networks ”. In: Thirty-sixth AAAI Conference on Artificial Intelligence (2022).

Summary: We design a tree search algorithm to look for sound reactive scheduling strategies for a hard scheduling problem known as Disjunctive Temporal Networks with Uncertainty. We leverage a Graph Neural Network (GNN) for guidance in the tree search for performance improvements. While the search is performed only in promising domain areas, which may lead to loss of coverage of existing strategies, experiments show that significant coverage of existing strategies is retained with proper time discretization. As a result, the algorithm almost always finds a reactive strategy when one exists, and does so significantly faster than the state-of-the-art solver even without using GNN guidance. When GNN guidance is provided,  the algorithm is able to solve significantly harder problems as well.

Osanlou, Kevin, Jeremy Frank, J. Benton, Andrei Bursuc, Christophe Guettier, Eric Jacopin, and Tristan Cazenave. “Time-based Dynamic Controllability of Disjunctive Temporal Networks with Uncertainty: A Tree Search Approach with Graph Neural Network Guidance”. In: Workshop on bridging the Gap Between AI Planning and Reinforcement Learning (PRL). International Conference on Automated Planning and Scheduling (ICAPS).

Osanlou, Kevin, Andrei Bursuc, Christophe Guettier, Tristan Cazenave, and Eric Jacopin. “Optimal Solving of Constrained Path-Planning Problems with Graph Convolutional Networks and Optimized Tree Search”. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, pp. 3519–3525.

Summary: We leverage a Graph Neural Network (GNN) for path-planning purposes in this work. The GNN is used to provide an early upper bound to a Branch & Bound tree search algorithm which looks for an optimal solution path. Results show the upper bound provided by the GNN allow significant branch cuts and enable the algorithm to outperform handcrafted heuristics and scale to problem complexity.

Osanlou, Kevin, Christophe Guettier, Andrei Bursuc, Tristan Cazenave, and Eric Jacopin . “Learning-based Preference Prediction for Constrained Multi-Criteria Path-Planning”. In: Scheduling and Planning Applications workshop (SPARK). International Conference on Automated Planning and Scheduling (ICAPS).

Summary: We combine machine learning models with constraint programming algorithms for multi-criterion path-planning with constraints. More specifically, we predict an unknown criterion and look for pareto-optimal solutions with a known criterion. Experiments show the approach yields good quality solutions for the unknown criterion at the cost of a small increase of the known criterion.

Osanlou, Kevin, Christophe Guettier, Andrei Bursuc, Tristan Cazenave, and Eric Jacopin . “Constrained Shortest Path Search with Graph Convolutional Neural Networks”. In: Workshop on Planning and Learning (PAL-18). International Joint Conference on Artificial Intelligence (IJCAI).

Summary: We combine a Graph Neural Network (GNN) with a constraint programming solver for path-planning under constraints. More specifically, after the GNN is trained, inference is performed so as to define optimal variable search order for the solver. Results exhibit significantly increased solving perfermance on multiple graphs.