Joëlle Pineau

Pineau’s education combines a BASc in engineering at Waterloo with an MSc and PhD in robotics at Carnegie Mellon. Before joining Cohere, she led Meta’s Fundamental AI Research organization; her academic biography records work there during 2017–2025. She remains a McGill professor and a core member of Mila, with research applications in robotics, healthcare and conversational agents.[1][2][6][7][8]

Plan from a manageable set of possible beliefs

With Geoff Gordon and Sebastian Thrun, Pineau introduced Point-Based Value Iteration for partially observable Markov decision processes. Instead of computing an exact solution across every possible belief about the world, it updates values at representative belief points and expands that set. The paper combines an error analysis with robot laser-tag and benchmark experiments, making planning under uncertain observations more tractable.[3]

Explain the family of point-based solvers

Pineau later coauthored a survey with Guy Shani and Robert Kaplow that organizes extensions to point-based planning. It compares how methods select belief points and order their updates, then evaluates them on shared benchmarks. That perspective helps a reader connect an individual algorithm to the broader choices involved in solving a partially observed planning problem.[5]

Make the experimental record easier to reproduce

The Machine Learning Reproducibility Checklist asks researchers to state assumptions, data splits, excluded data, dependencies, training and evaluation commands, hyperparameter selection and variability across runs. Its accompanying guidance separates paper and code artifacts, making it useful when preparing a submission as well as when attempting to repeat a result.[4]