miguelgondu's blog
About Me
I'm Miguel González Duque, a mathematician from the Universidad Nacional de Colombia. I'm mostly interested in the intersection between probabilistic modeling and geometry, and their applications to real-world problems. I'm also really curious about Python development, as can be seen in this blog and in my GitHub profile.
I have worked on really diverse projects. From building web applications and APIs for content adaption, through learning latent representations in robotics and video games (see below), to learning the different topics that are discussed in Colombia's longest running philosophy journal automatically using LDAs.
I'm currently a PhD student at the Creative AI Lab at the IT University of Copenhagen, under the supervision of Prof. Sebastian Risi. We are developing systems that create and adapt game content to users. I have also collaborated with Søren Hauberg's lab at DTU, working on generative modelling and differential geometry. I did a 6-month PhD Sabbatical at the Bosch Center for AI working with Leonel Rozo on using Gaussian Processes for representation and regression of robotic data.
For more details on my professional experience, check my LinkedIn.
Feel free to contact me at miguelgondu(at)gmail(dot)com.
Some publications
For a complete list of publications, check my Google Scholar profile.
MarioGPT: Open-ended Text2Level Generation through Large Language Models
Shyam Sudhakaran, Miguel González-Duque, Claire Glanois, Matthias Freiberger, Elias Najarro, Sebastian Risi
Shyam fine-tuned GPT2 with a few hand-labeled examples of Super Mario Bros levels. Matthias and I analysed how different its paths were from an A-star agent's.
Bringing robotics taxonomies to continuous domains via GPLVM on hyperbolic manifolds
Noémie Jaquier, Leonel Rozo, Miguel González-Duque, Viacheslav Borovitskiy, Tamim Asfour
We embed humanoid poses in hyperbolic space, and have the latent distances match a robotics taxonomy. Main work done by Noémie Jaquier.
Mario Plays on a Manifold
Miguel González-Duque, Rasmus Berg Palm, Søren Hauberg, Sebastian Risi.
Presented at CoG 2022We train 2-dimensional VAEs on Mario and Zelda and define a graph in latent space only where playable levels live. This lets us sample and interpolate safely.
Variational Neural Cellular Automata
Rasmus Berg Palm, Miguel González-Duque, Shyam Sudhakaran, Sebastian Risi.
Presented at ICLR 2022We used an NCA as the decoder of a Variational Autoencoder. Main work done by Rasmus Berg Palm.
arxiv - publication - code
Pulling back information geometry
Georgios Arvanitidis, Miguel González-Duque, Alison Pouplin, Dimitrios Kalatzis, Søren Hauberg
Presented at AISTATS 2022We induced geometries in latent space by pulling back the Fisher-Rao metric.
Fast Game Content Adaptation Through Bayesian-based Player Modelling
Miguel González-Duque, Rasmus Berg Palm, Sebastian Risi.
Presented at CoG 2021We applied a Bayesian Optimization approach to dynamically adjusting difficulty.
arxiv - publication - code
Finding Game Levels with the Right Difficulty in a Few Trials through Intelligent Trial-and-Error
Miguel González-Duque, Rasmus Berg Palm, David Ha, Sebastian Risi.
Presented at CoG 2020We used Gaussian Processes and Bayesian Optimization to find good levels for one bot using a prior from another bot.
arxiv - publication - code
Learning a Behavioral Repertoire from Demonstrations
Niels Justesen, Miguel González-Duque, Daniel Cabarcas Jaramillo, Jean-Baptiste Mouret, Sebastian Risi
Presented at CoG 2020We conditioned a StarCraft2 bot on strategies and optimized for the best strategy. This image was made by Niels Justesen.
Neural Networks that express multiple strategies in the video game StarCraft 2
Miguel González-Duque.
My M.Sc. Thesis (Mathematics at UNAL-Med). During my first stay at ITU, I collaborated with Niels, Sebastian and Daniel Cabarcas on applying dimensionality reduction to strategies, and creating bots for StarCraft 2.