NNRG Spotlights

DISLEX Model tested in a Clinical Trial

A new NIH grant, Predicting Rehabilitation Outcomes In Bilingual Aphasia Using Computational Modeling started in July 2016 with familiar personnel: Uli Grasemann returned to UT to lead the computational modeling work, and Swathi Kiran at Boston University leads the clinical work. Building on our prior grant, the goal is to use self-organizing maps with associative connections, i.e. the DISLEX model, to understand damage and recovery of the lexical system in bilinguals. However in this new grant (starting this summer), the model is put into an actual clinical trial, predicting optimal treatment for actual patients. To our knowledge, this is the first time a computational cognitive model has served in this role! See our earlier paper for a background on this project.

Previous Spotlights (hide)

Evolutionary Computation in the Real World

Sentient Technologies, Inc. (where Risto currently is on a leave) just launched a novel application of evolutionary computation on e-commerce. Their Ascend is a system for designing web pages automatically in order to maximize conversions (i.e. user actions such as purchases or sign-ups). Evolution generates web page candidates that are then tested on real users in real time. This blog post describes the technology in more detail.

NNRG in the media
8/22/2015 (updated 3/15/2016)

The work of NNRG and its alumni has recently been featured in a number of media outlets. For instance, UT's press release on Joel Lehman and Risto Miikkulainen's PLOS ONE article on how extinctions can accelerate evolution was picked up by a number of sites, including the National Science Foundation, Science Daily, phys.org, and even the Daily Mail. It generated much discussion e.g. on Twitter and io9, and a hilarious interview/skit with TexasExes. Ken Stanley and Joel Lehman's new book on novelty search was the focus of an article on FiveThirtyEight and Ken's radio interview. Dan Lessin's research was featured in the Robohub robotics news site. Risto was recently interviewed about game AI by the KXAN TV channel and about strong AI by the Austin American Statesman newspaper, and about evolutionary computation by David Beyer for O'Reilly Radar Podcast. Also, a well-known gamer Seth Bling used NEAT to create a Mario-playing agent---a YouTube video of it has over 2 million views. The word is getting out!

Schrum wins GECCO Best Paper Award---again!

Jacob Schrum and Risto Miikkulainen won the Best Paper Award in the Digital Entertainment and Arts track at the GECCO 2015 Conference for their paper on Solving Interleaved and Blended Sequential Decision-Making Problems through Modular Neuroevolution. Jacob received his PhD in Spring 2014 and is currently an Assistant Professor at Southwestern University. This is his second such award, following the one from GECCO-2014. In addition he has won two best paper awards from the IEEE Conference on Computational Intelligence and Games, in 2011 and 2009. Way to go Jacob!

Schrum wins GECCO Best Paper Award

Jacob Schrum and Risto Miikkulainen won the Best Paper Award in the Digital Entertainment and Arts track at the GECCO 2014 Conference for their paper on Evolving Multimodal Behavior With Modular Neural Networks in Ms. Pac-Man. Jacob received his PhD in Spring 2014 and will be an Assistant Professor at Southwestern University in Fall 2014.

NIH and IARPA projects started

In the Role of Emotion and Communication in Cooperative Behavior project the goal is to understand how these cognitive facilities contribute to mobbing behavior, as observed in hyenas and in a computational model. The project is funded by the Modeling Social Behavior of the NIH. In the Knowledge Representation as Embodied Abstractionsproject a team of researchers (led by Teledyne Scientific, Inc., and including the NNRG at UT) is aiming to model how representations of concepts in the brain are shaped by context. This project is funded by the Knowledge Representation in Neural Systems program of IARPA.

Lockett and Waters win awards

Alan Lockett received a runner-up award for the Best Overall Paper at the 2013 IEEE Congress on Evolutionary Computation for his paper on Measure-Theoretic Analysis of Performance in Evolutionary Algorithms. Alan received his PhD with the NNRG last year and is currently a postdoctoral researcher at IDSIA, Switzerland.

Austin Waters and Risto Miikkulainen received an IAAI Deployed Application Award at the 2013 AAAI Conference for their paper on GRADE: Machine Learning Support for Graduate Admissions. GRADE is used in the UTCS Department to evaluate applicants to the PhD program, reducing the workload of the admissions committee by 74%.

Neuroevolution Scholarpedia Article Published

Lehman and Miikkulainen recently authored an article on Neuroevolution in Scholarpedia, the peer-reviewed open-access encyclopedia. This article reviews more than two decades of research on evolution of neural networks, a technique that has proven to be a practical engineering methodology as well as a novel tool for exploring biological hypotheses. This article is the the latest step in our effort to educate people in this area, following several tutorials since 2005 at GECCO, IJCNN, WCCI, and Alife, and the development of OpenNERO software platform for AI education.

The UT^2 Bot wins BotPrize

For the first time in the five-year history of the BotPrize Competition (i.e. the "Turing test for game bots" in the Unreal 2004 video game), the best bots were judged as human more than 50% of the time, resulting in the awarding of the actual BotPrize sponsored by 2K Games. Two bots achieved this feat; one of them was the UT^2 bot created by Igor Karpov, Jacob Schrum, and Risto Miikkulainen, based on neuroevolution of combat behavior and navigation based on human traces. For more details, see our BotPrize project page and the official BotPrize Competition page, and listen to/watch interviews on Digital Nibbles and aigamedev.org.

The UT^2 Bot wins Humanlike-Bot Competition

As part of the IEEE World Congress on Computational Intelligence (WCCI 2012) in Brisbane, Australia, a competition was held in a "Turing test for game bots", i.e. for the most humanlike game agents in the Unreal 2004 video game. The UT^2 bot, created by Igor Karpov, Jacob Schrum, and Risto Miikkulainen won the competition with a 21% humanness rating, which for the first time was better than half the human players. See this year's results and general description of the BotPrize Competition for more details.

Online OpenNERO Tournament Demonstrates a "Machine Learning Game"

The tournament was run in the NERO machine learning game of the OpenNERO software system in Dec 2012. Most of the 159 participants came from the Stanford Online AI course, and some from the cs343 and cs394n courses at UT Austin. Most teams were trained with neuroevolution and some with Q-learning. The results showed that surprisingly diverse strategies can be successful, and no single strategy dominates all others---which indeed makes the game an interesting challenge for humans and machine learning alike. See the tournament results page for details.

Schrum wins CIG-11 Best Paper Award

Jacob Schrum won the Best Paper Award at the 2011 IEEE Conference on Computational Intelligence and Games for his paper with Risto Miikkulainen on Evolving Multimodal Networks for Multitask Games. Multitask games are domains in which separate tasks within the domain each have their own dynamics and objectives. The methods described in the paper allow evolved neural networks to use multiple output modes to better learn multiple modes of behavior in multitask games. Videos of evolved behavior can be seen here.

Silverthorn again wins solver competition

The Borg algorithm portfolio, developed by Bryan Silverthorn, has placed first in the main category of the annual pseudo-Boolean solver competition for the second year in a row. The competition tests a program's ability to solve difficult computational problems, such as finding bugs in software and hardware, expressed in "pseudo-Boolean" mathematical form. Borg succeeds by using multiple algorithms and automatically choosing which to apply to each problem. Its recent improvements include better predictions of whether an algorithm will succeed, and better schedules of when an algorithm should be executed. Borg's continued success demonstrates the promise of its portfolio approach: solving larger sets of more challenging problems, more quickly.

Computational Modeling of Schizophrenia

Uli Grasemann's work (with Ralph Hoffman and Risto Miikkulainen) on computational modeling of schizophrenia has reached an unusually wide audience. After UT did a press release on it, the story was Slashdotted, and then spread to several venues from the Chronicle of Higher Education and NPR's Science Friday to somewhat unexpected venues such as IEEE Spectrum, Business Week and Popular Mechanics. Some of the most interesting headlines include "Computer takes responsibility for terrorist bombing" and "Boffins develop method of driving computers insane." Nevertheless, the interest demonstrates that connectionist modeling has perhaps become mature enough so that it can be now a topic of public discussion.

BEACON Center started

BEACON is an NSF Science and Technology Center, focusing on "Evolution in Action", i.e. interdisciplinary approaches to understanding biological and computational evolution. It is a consortium of Michigan State University (the lead institution), University of Idaho, University of Texas, University of Washington, and North Carolina A&T University, with a total funding of $25M over five years. Several people in the NNRG participate in BEACON's activities, and Risto is the lead for the UT subcontract. For more details, see the BEACON website.

Silverthorn wins solver competition

Bryan Silverthorn's Borg program won the main category of the Fifth Annual Competition of Pseudo-Boolean Solvers at the Thirteenth International Conference on Theory and Applications of Satisfiability Testing. This international competition compared systems for solving pseudo-Boolean expressions, which are used to address problems such as the correctness of electronic circuit designs. Bryan's program succeeded by applying multiple solution methods, executing them according to statistical patterns that it automatically discovered from experience. For more details see the Borg project page.

Three projects started, with funding from NSF and Google

In the NSF Robust Intelligence project, our work on evolving neural networks in control tasks will be extended to sequential decision making at the strategic level. It makes use of the OpenNERO project, funded by Google, Inc., where we are extending the NERO machine learning game into a general AI research and education platform. In the interdisciplinary NSF SciSIP project, the goal is to understand competitive multiagent search, and apply it to organizational theory on how high-technology firms can best innovate.

Schrum wins CIG-09 Best Student Paper Award

Jacob Schrum won the Best Student Paper Award at the 2009 IEEE Symposium on Computational Intelligence and Games for his paper with Risto Miikkulainen on Evolving Multi-modal Behavior in NPCs. See also videos of evolved behaviors.

New NSF, NIH, and ARP projects started

In the NSF-supported "CreativeIT" project, the goal is to leverage human intuition with neuroevolution discovery of complex behavior. In the ARP project, we are utilizing simulated environments like OpenNERO to learn complex high-level behaviors. Both of these projects build on our prior work on evolving neural networks. The NIH project, in turn, continues our cognitive science research stream: A model of bilingual lexicon is being built in order to study how it breaks down in aphasia, and how the performance can be best regained through rehabilitation.

Valsalam and Miikkulainen win a GECCO Best Paper Award

Vinod Valsalam and Risto Miikkulainen won the Best Paper Award in the "Artificial Life, Evolutionary Robotics, Adaptive Behavior, Evolvable Hardware" track of GECCO-2008 for their paper Modular Neuroevolution for Multilegged Locomotion. See also videos of walking behaviors.

New NNRG website launched

The site has moved to a departmental server, it has a new look and feel (under the same design), and now includes a "demos" area that will be populated over the next few months. Please bear with us as we complete the transition and add the content.

NERO 2.0 released

A new, significantly extended version of the NERO Machine Learning Game, based on rtNEAT neuroevolution, was released in Fall 2007. This version includes an interactive game mode, a new user interface, and more extensive training tools.

Reisinger and Miikkulainen win a Best Paper Award

Joe Reisinger and Risto Miikkulainen won the Best Paper award in the Generative and Developmental Systems track at GECCO 2007, for their paper "Acquiring Evolvability through Adaptive Representations."

Computational Maps in the Visual Cortex published

This book, co-authored by Risto Miikkulainen, James A. Bednar, Yoonsuck Choe, and Joseph Sirosh, and published by Springer, reviews 18 years of research by the UTCS Neural Networks Research Group on modeling the visual cortex. The book is accompanied by the Topographica software package for general modeling of computational maps.

Bednar, Grasemann, Miikkulainen, Sit, and Valsalam win awards at GECCO 2005

Valsalam, Bednar, and Miikkulainen's Constructing Good Learners using Evolved Pattern Generators won the Best Paper Award in the Evolutionary Robotics, A-life, and Adaptive Behavior track, and Sit and Miikkulainen's Learning Basic Navigation for Personal Satellite Assistant using Neuroevolution in the Real World Applications track; Grasemann and Miikkulainen's Effective Image Compression using Evolved Wavelets won a Bronze Medal in the Human-Competitive Results Competition and was featured in NSF Discoveries.

Stanley, Bryant, and Miikkulainen win Best Paper Award

Ken Stanley, Bobby Bryant, and Risto Miikkulainen won the Best Paper Award at the IEEE 2005 Symposium on Computational Intelligence and Games, for Evolving Neural Network Agents in the NERO Video Game. Risto gave a keynote talk and NERO producer Aliza Gold presented a paper on transfering academic AI to game applications.

NERO: A Pioneering Neuroevolution-based Video Game

NERO demonstrates how neuroevolution can be used to adapt behaviors of autonomous agents in a complex real-time video game. Together with a team at UT Digital Media Collaboratory, Ken Stanley and Bobby Bryant have now completed the first prototype of NERO.

Neuroevolution Researchers Address Game Development Workshop

Members of our neuroevolution group join stars from the world of computer game design in the second annual Game Development Workshop organized by the UT Digital Media Collaboratory, in an effort toward establlishing an ongoing technology transfer between academic AI research and the real-world considerations of commercial game development.

Gomez & Miikkulainen Win Best Paper Award

Faustino Gomez and Risto Miikkulainen won the "Best of GECCO" award in the Real World Applications track at GECCO 2003, for their paper "Active Guidance for a Finless Rocket Using Neuroevolution"

UT Neural Networks Research Group Web Site Goes Live

The Neural Networks Research Group at the University of Texas at Austin has redesigned and reprogrammed their website... and you're looking at it.

Stanley & Miikkulainen Win Best Paper Award

Ken Stanley and Risto Miikkulainen's paper, Efficient Reinforcement Learning Through Evolving Neural Network Topologies," is the winner of the 2002 Best Paper Award in Genetic Algorithms.