Jake Tuero

Jake Tuero

PhD student at the University of Alberta.

Research Interests: Machine learning, reinforcement learning, and tree search methods.

 

About Me

I’m currently a PhD student attending the University of Alberta, under supervision of Michael Buro and Levi Lelis. My research is focused on creating agents which can solve complex tasks with sparse rewards, that requires a well thought out sequence of high level actions to solve. Specifically, environments in which high level plans may be correct, but the implementation of each step impacts whether the rest of the plan is still achievable. I did my undergrad in Computer Science and Mathematics at Wilfrid Laurier University.

CV

You can find an up to date copy of my CV here.

Projects

For a complete description and walkthrough of my projects, see here.

  • MuZero-CPP: A pure C++ implementation of MuZero, using libtorch C++, designed for flexibility and speed. Implemented features includes multi-threaded async actor inference, complex action representation, and fast batched inference on the GPU.
  • ptutil: Personal PyTorch framework library for common boilerplate code, used in my thesis work and personal projects.
  • Stones n Gems: Author of the game Stones n Gems for the Open Spiel framework by DeepMind. Stones n Gems is a simplified version of a mixture of common stone and gem games, such as Boulder Dash and Emerald Mines.
  • Rocks n Diamonds: A wrapper for the Rocks’n’Diamonds open source C arcade style game (based off Boulder Dash, Emerald Mines, Supapplex, and Sokoban). The project extends Rocks’n’Diamonds by letting users add their own AI controllers, providing a host of library functions to easily access the internal state of the engine, replay functionality, and logging.

Research

For a complete description and walkthrough of my research, see here, or my Google Scholar page.

  • Bayes DistNet - A Robust Neural Network for Algorithm Runtime Distribution Predictions: We extend runtime distribution prediction models into the Bayesian setting for the first time, using the current state-of-the-art model as a baseline. We show that our model achieves robust performance in the low observation setting, as well as handling censored observations. Our model can also quantify its uncertainty in its predictions. Paper was accepted to AAAI-21.

  • Chromatic Symmetric Functions and H-free Graphs: We consider the e-positivity question for H-free graphs, where H = {claw, F}, F is a four-vertex graph. We settle the question for all cases, except H = {claw, co-diamond}, and we provide some partial results in that case. Paper was accepted to Graphs and Combinatorics - Springer.