Researcher Engineear @ InstaDeep
Full-time, Sep. 2019 - present.
I am working on the development of our reinforcement-learning-based decision-making product.
Part-time, Oct. 2018 - Sep. 2019.
I participated in the development of our reinforcement-learning-based decision-making product. I worked on the neural network design and experimental comparisons with mainstream reinforcement learning algorithms (PPO, AlphaZero, etc.). Previously, I developed a novel model-free algorithm which outperformed plain Monte-Carlo tree search (MCTS).
During 2018, my principal project was to extend the research results of Ranked Reward to glass cutting optimization problem, an industrial combinatorial optimization problem. Among all participants, our solution was the only one which did not depend on traditional optimization techniques and we ranked 16 out of 60 teams.
Internship, Feb. 2018 - Sep. 2018.
My main project was to use reinforcement learning to solve the bin packing problem, which is an NP-hard combinatorial problem. The objective of bin packing is to place boxes in containers efficiently, i.e. minimizing the waste of space, while respecting multiple physical constraints.
Based on AlphaZero, a reinforcement learning algorithm designed for two-player games, we proposed Ranked Reward to enable self-playing training in single-player games. Ranked Reward reshapes the rewards based on the agent’s previous performances to create a relative metric. Without specific human knowledge, the proposed method achieves super-human performance and surpasses MCTS, supervised agent, heuristic algorithm, and commercial integer programming solver (Gurobi) on both 2D and 3D bin packing problems.
This work has been published (arXiv) as workshop paper in both NeurIPS and AAAI. I also received the research internship award from École Polytechnique, France.
Besides the bin packing project, I participated in some other research projects as well, including evolutionary algorithms and auto machine learning. Specially, we presented our latest results on neural network architecture search at Deep Learning Indaba 2018.
Operations Research Analyst @ SNCF Réseau
Internship, Jun. 2017 - Sep. 2017.
- Modelled railway scheduling as a multi-objective optimization problem.
- Solved the scheduling problem with discrete programming (local search, tabu search, and simulated annealing).
- Outperformed human experts and approached the performance of LocalSolver.
More unlabelled data or label more data? A study on semi-supervised laparoscopic image segmentation
Yunguan Fu, Maria R. Robu, Bongjin Koo, Crispin Schneider, Stijn van Laarhoven, Danail Stoyanov, Brian Davidson, Matthew J. Clarkson, Yipeng Hu
Accepted to MICCAI 2019 Medical Image Learning with Less Labels and Imperfect Data Workshop, arXiv
- Implemented image segmentation using U-Net and semi-supervised Mean Teacher model for laparoscopic images.
- Demonstrated that the specific training method of Mean Teacher is responsible for the performance improvement besides the additional unlabeled data.
- Demonstrated that adding more unlabeled data potentially could provide similar performance improvement compared to using more labeled data.
Ranked Reward: Enabling Self-Play Reinforcement Learning for Combinatorial Optimization
Alexandre Laterre, Yunguan Fu, Mohamed Khalil Jabri, Alain-Sam Cohen, David Kas, Karl Hajjar, Hui Chen, Torbjorn S. Dahl, Amine Kerkeni, and Karim Beguir.
Accepted to NeurIPS 2018 Deep RL Workshop and AAAI 2019 RL in Games Workshop, arXiv.
- Implemented AlphaZero, R2, and heuristic algorithm to solve bin packing problems.
- Designed neural network architectures to solve both 2D and 3D problems.
- Designed a supervised dataset to evaluate the neural network architectures.
- Designed and implemented linear programming models for bin packing problems using Gurobi solver.
- Conducted experiments to compare with MCTS, heuristic, supervised agent and Gurobi solver.