Stefano Ermon

Stefano Ermon

Assistant Professor, Department of Computer Science
Fellow, Woods Institute for the Environment
Stanford University

Office: Gates Building #228

Phone: (650) 498-9942

Email: ermon AT

Research Group Website

About Me

I am an Assistant Professor in the Department of Computer Science at Stanford University, where I am affiliated with the Artificial Intelligence Laboratory and a fellow of the Woods Institute for the Environment.

My research is centered on techniques for scalable and accurate inference in graphical models, statistical modeling of data, large-scale combinatorial optimization, and robust decision making under uncertainty, and is motivated by a range of applications, in particular ones in the emerging field of computational sustainability.





Honors and Awards


Professional Service


Software and Benchmarks

WISH is a general algorithm to approximate (with high probability and within any desired degree of accuracy) discrete weighted sums defined over exponentially large sets of items. This implementation is specifically designed to approximate the partition function (normalization constant) of discrete probabilistic graphical models, by (approximately) solving a small number of optimization instances (maximum likelihood queries) using a combinatorial optimization package.
  • Source [UAI 2015]
  • Binaries (Linux, 64 bit) of WISH version 1.0 [ICML 2013]
  • Source [UAI 2013]
  • Instances used to evaluate the WISH algorithm [ICML 2013]
MoNet is an algorithm to infer latent network structure based on observations of textual "cascades" spreading over the network. For example, MoNet can be used to infer a following relationship (who is following whom) on Twitter by observing sequences of tweets. It first defines a probabilistic model that specifically takes into account time and textual information of the messages, and then infers the most likely underying network structure by solving a sequence of convex optimization problems.
  • MoNet source code for inferring latent network structure based on textual cascades [ECML-PKDD 2012]

Combinatorial Materials Discovery Benchmark Instances
In combinatorial materials discovery, materials scientists search for new materials with desirable physical properties by obtaining x-ray diffraction measurements on hundreds of samples from a composition spread. We integrated domain-specific scientific background knowledge about the physical and chemical properties of the materials into a Satisfiability Modulo Theories (SMT) reasoning framework based on linear arithmetic. Using a novel encoding, state-of-the-art SMT solvers can automatically analyze large synthetic datasets, and generate interpretations that are physically meaningful and very accurate, even in the presence of noise.
  • SMT instances (SMTLib Version 2.0) for the phase map identification of the Al-Li-Fe synthetic system [SAT 2012]

Search Tree Sampler
SearchTreeSampler is a sampling technique that, while enforcing an approximately uniform exploration of the search space, leverages the reasoning power of a systematic constraint solver in a black-box scheme. They key idea is to explore the search tree uniformly in a breadth-first way, subsampling a subset of representative nodes at each level. The number of nodes kept at each level is a parameter used to trade off uniformity with computational complexity. The samples provided by STS can then be used to estimate the number of solutions of the problem (partition function).
  • Source code of Search Tree Sampler (with MiniSAT as an NP-oracle). [UAI 2012]

Energy Demand in Commuter Trips Dataset
We produced a dataset of energy demand profiles for commuter trips acros the US by processing the raw data originally collected by the ChargeCar project at CMU. We used this dataset to train an intelligent energy management system for electric vehicles, based on a combination of optimization, MDPs, and supervised learning techniques. The new approach significantly outperforms the leading algorithms that were previously proposed as part of an open algorithmic challenge.
  • Dataset of crowdsourced commuter trips acroos the US (generated processing the raw data originally collected by ChargeCar) [ECML 2012]

Flat Histogram Sampling Code
We investigated the use of advanced flat histogram sampling techniques from statistical physics to explore large combinatorial spaces. This is a class of adaptive MCMC (Markov Chain Monte Carlo) methods that can adapt transition probabilities based on the chain history. Intuitively, this allows the chain to escape from local minima, which can be very helpful for difficult energy landscapes. Further, we introduced a focused component (inspired by local search combinatorial optimization) that can leverage problem structure and speed up convergence.



You can find out more about me and see some pictures here.


©2017 Stefano Ermon