GAIA: Adaptive Game-Playing Agents

The designs of long-living interactive games evolve through many versions. Changes from one version of a game to the next are typically incremental and often very small. A game designer (or a team of game designers and software engineers) formulates the requirements of the new version of the game, adapts the software for playing the previous versions to meet the new requirements, and implements and evaluates the modified designs of the game and the software. Typically the game designer uses high-level scripting languages to define the game environment (e.g. percepts, actions, rules, constraints) as well as the behaviors of various virtual agents in the game.

We posit that an interesting research issue in game playing is how a virtual agent might adapt its design, and thus its behaviors, to very small changes in its game environment. If the changes in the game environment can be arbitrarily large and complex then this becomes an “AI-complete problem.” However, even if the changes to the game environment are incremental and very small, this is a hard computational problem because changes to the environment can be of many types, modifications to the agent design can be of many types, there is no one-to-one mapping between changes to the environment and modifications to the agent design, and any modification to the agent design needs to be propagated down to the level of program code so that the new software is directly executable in the game environment.

The goal of this project is to develop an interactive environment called GAIA (for Game Agent Interactive Adaptation) in which the game designer generates requirements for a new version of a game, and the designer and the legacy software agents from previous versions of the game cooperatively adapt the agent designs to the new game requirements. We are developing and testing our meta-reasoning technique for adapting a mature program in the domain of turn-based, multi-player, strategy games (specifically FreeCiv).



How to Work at Home and Get Stuff Done

March 18, 2020

How to Work at Home and Get Stuff Done

Eric Gregori

School of Interactive Computing, Georgia Institute of Technology,

A Little About Myself for Reference

I am 49 years old with a wife, 2 teenage boys (1 driving), and 2 dogs. I have worked as either a software engineer or an engineering manager for 25 years of which 17 of those years have been work at home. I earned a master’s of computer science from the Georgia Institute of Technology’s OMSCS online learning program. I currently work remotely as a research scientist for Georgia Tech.

Working From Home is Very Different From Working in an Office

Whether you sit in front of a computer at home or you sit in front of a computer in an office the actual work you are doing is the same. At the task level, the job does not change. This is a very important concept to remember; you are paid to complete tasks, not work hours. When working at home the environment changes, communications change, but the tasks (and the need to complete tasks) stay the same. Unfortunately, your home is comfortable, full of distraction, and a barrier to communications with your manager and coworkers. All of which work against you when you are trying to get work tasks (stuff) done.

Your Home has Been Designed for Comfort, not Productivity

The first obstacles to getting stuff done at home are the couch, TV, and refrigerator. These distractions were purchased for comfort, to wind down after a day or week of hard work. These items will call to you when you are trying to get stuff done. The solution is a workspace out of sight of those distractions; preferably, a dedicated workspace with a door [1]. When it is time to work, go to your workspace and if you have a door close it. Your home workspace should be designed for productivity by following the same rules as an office workspace in terms of layout and amenities.

Human Distractions

I have a wife and two teenage boys that I love very much, but they keep me from getting stuff done by assuming that if I am home I am available. If you have a door in your workspace, close it, and when you are on a conference call put a sign on the door stating that you are on a video call and are not to be disturbed. Invest in a pair of comfortable noise-canceling headphones with a microphone. If you do not have a door, explain to your family that at any time you may be on a conference call and to approach you assuming you are on a call. I also use hand signals to signal to my family that I am on a call.

Out of Sight, Out of Mind – Communication is key

In my experience working at home as a manager and an individual contributor, this proverb is just applicable now as it was in 850 B.C. How do you get stuff done if you are not sure what the stuff is or you are unsure of the definition of “done”. Most people do not realize the amount of important information that is conveyed via impromptu hallway meetings until they have worked at home for a month or two. “Out of sight, out of mind” is human nature [3] and must be actively counteracted in order to understand what stuff needs to be done.

Over-communicating with your manager, your teammates, and your internal customers is the solution to, “out of sight, out of mind”. Check your email often during your manager’s business hours, be available on Slack, strive to respond to communications immediately. This is very easy to do by installing email and Slack on your phone. I consider it an important part of my job to keep my manager and team informed about what is going on, so I tend to send a lot of emails. This is very dependent on your manager; some of my past manager’s liked the fact that I kept them in the loop via email and others would rather be updated during conference calls. Make sure you figure out how your manager likes to communicate. Even if you have a manager that would rather be updated via a call, send an agenda email before the call with your update. In my experience, you are always better off over-communicating than under communicating when working at home.

Career Impact

Although over-communicating can do a lot to combat “out of sight, out of mind”, research has shown there is a potential work from home promotion (WFH) ‘discrimination’ penalty [2]. This is situation-specific in my experience as both a manager and an individual contributor. Working at home needs to be factored into your career goals with the understanding that you may not have the same career advancement options as someone who works in the office.

The Advantages of Working at Home

If done right, research has shown that working from home (WFH) is beneficial to both the employee and the employer [2]. Success requires a unique manager and employee who are both invested in making WFH work. The manager must be willing to trust and give up a little control, while the individual contributor must be willing to spend more time communicating.

The reward for the company is a more productive employee; research shows WFH employees are more productive than office employees [1]. The reward for the WFH employee is flexible hours, no commute, and a flexible work environment. For example, I enjoy working at night so I split my day between meetings during business hours and coding at night; sleep flexibility is another advantage of working from home.


[1] Crosbie, Tracey, and Jeanne Moore. “Work–life balance and working from home.” Social Policy and Society 3.3 (2004): 223-233.

[2] Bloom, Nicholas, et al. “Does working from homework? Evidence from a Chinese experiment.” The Quarterly Journal of Economics 130.1 (2015): 165-218.

[3] Schindler, Andreas, and Andreas Bartels. “Parietal cortex codes for egocentric space beyond the field of view.” Current Biology 23.2 (2013): 177-182.

Author Biography: Eric Gregori is a Research Scientist in the Design Intelligence Lab at Georgia Institute of Technology. His research interest revolves around the use of Knowledge Engineering, Machine Learning, and Natural Language Understanding in education and customer support. He is currently working on the Jill Watson conversation agent to support education by amplifying the utility of teachers. He holds an MS in computer science from Georgia Tech.

Lessons learned from teaching an online course in Georgia Tech’s OMCSCS program

March 17, 2020

Online Pedagogy: Lessons learned from teaching an online course in Georgia Tech’s OMCSCS program

Ashok K. Goel ( Design & Intelligence Laboratory, School of Interactive Computing, Georgia Institute of Technology


The spread of COVID-19 has led to an unprecedented move towards online education across the country and around the world. It is too early to predict the long-term impact of this extraordinary move: it may be temporary and much of education may soon revert back to the old normal of in-person classes, or it may lead to a large-scale shift towards online learning. Even if the move is only temporary, it may lead to a new normal of blended learning in which online educational materials, home works and assessments support in-person learning. In any case, this raises fundamental questions for online pedagogy: How do we develop a successful online course? How do we prepare high quality educational materials for online classes?

In 2014, my Georgia Tech colleague David Joyner and I faced exactly these questions when we developed an online course on Knowledge-Based Artificial Intelligence (KBAI) as part of Georgia Tech’s Online Masters of Science in Computer Science program (OMSCS; We identified about 150 KBAI concepts, methods and skills we wanted students to learn, developed a set of 26 video lessons, and designed a suite of learning assessments including home works, design and programming projects, and take home examinations (Goel & Joyner 2016, 2017). In Fall 2014, we offered the online class to about 200 students in the OMSCS program. In Fall 2015, I transformed the pedagogy I had been using in the in-person KBAI class for graduate and undergraduate residential students for more than a decade into blended learning (Goel 2019). In Spring 2016, my Design & Intelligence research laboratory developed a virtual teaching assistant called Jill Watson for automatically answering routine questions on the discussion forum of the online KBAI class (Goel & Polepeddi 2017). We estimate that more than 6000 students have taken the online and blended versions of the KBAI class since Fall 2014 and more than 150 (human) teaching assistants have helped with the teaching and learning in the class.

Assessing the quality of learning in any class is a complex matter, whether the class is online, blended or in-person. Three types of data indicate that the quality of learning in the online KBAI class is comparable to that in the in-person class for residential students (Goel & Joyner 2016, 2017; Goel 2019). First, surveys of online students from Fall 2014 through Fall 2019 report the same kind and degree of satisfaction with the online KBAI class as do the residential students with the in-person KBAI class during the same period. Second, the completion ratio in the online KBAI classes during this period has been comparable to that in the in-person KBAI classes, which is atypical for online classes. Third, the performance of the online KBAI students on the same set of learning assessments has been comparable to that of the residential students in the in-person KBAI classes. To control for the difference in the student demographics between the online and residential students, in Fall 2018 and Fall 2019, we repeated the above quasi-experimental studies in online and in-person sections of the KBAI class offered only to residential students, and found similar results.

Design Principles for Online Classes

In Goel & Joyner (2016), we analyze the design principles for developing online classes in detail. Table 1 summarizes the 10 design principles underlying the online and blended KBAI classes. For example, the first principle suggests that the course instructor may want to explicitly the learning goals, outcomes, strategies and assessments before developing the online class. The eighth principle indicates that design of an online class is an iterative process, with each iteration based on feedback and reflection on the preceding iteration. Thus, it is important to collect and analyze feedback and leave time for deliberation and reflection.

Table 1: Design Principles for Online Classes (adapted from Goel & Joyner 2016)

  • Establish learning goals, outcomes, strategies, and assessments first
  • Allocate adequate time for design, development, and delivery
  • Deliberately recreate natural features of the residential class
  • Leverage the advantages of digital media for online learning
  • Design project-based learning carefully
  • Understand the audience
  • Break the isolation experienced by many online students
  • Solicit feedback and be ready to iterate
  • Leverage peer feedback and autograding wisely
  • Use the online class to enhance the residential class

Design Principles for Video Lessons

The 26 video lessons for the online and blended versions of the KBAI class embed about 150 exercises, one for each concept in the concept inventory, as well as about 100 tutors that provided adaptive feedback on many of the exercises (–ud409). My colleague Chaohua Ou with Georgia Tech’s Center for Teaching and Learning has analyzed the design of the videos, and student responses to them over several semesters, in detail (Ou, Joyner & Goel 2019). Table 2 summarizes the seven principles for designing video lessons for online classes derived from the analysis. For example, the seventh principle suggests the use of prepared visuals rather than drawing them at run time.

Table 2: Design Principles for Video Lessons (adapted from Ou, Joyner & Goel 2019)

  • Learning by example
  • Learning by doing
  • Adaptive feedback
  • Learning through reflection
  • Four-phase instruction principle (activation, demonstration, application, integration)
  • Personalization principle (visible instructors, conversational presence, on-screen coaches)
  • Multimedia principle (prepared visuals)


It is important to note three qualifications here. First, developing a successful online class requires expertise not only in the subject of the course, but also in information technology and learning science. It also requires a strong team, and significant financial and technological support. We were very fortunate to have all these assets in developing the online KBAI class in 2014. My research laboratory conducts research into KBAI and thus we had the expertise in the subject; we are computer scientists and thus we had some expertise in information technology as well; and Joyner and I already were developing expertise in learning science as part of his Ph.D. work. In addition, Georgia Tech generously provided financial and technological support for developing the video lessons for the online KBAI class.

Second, if a teacher’s goal is simply to convert the educational materials prepared for an in-person class to an online forum due to the spread of COVID-19 and the consequent need for social distancing, then we expect the design principles in Table 1 should be of immediate value. The principles in both Tables 1 and 2 are likely to add more value if in the longer term a teacher wants to develop a new online course from the start.

Third, learning in general is situated in the external world and thus is context dependent. The effectiveness and efficiency of learning depends not only on the individual learner, teacher, and educational materials, but also on the physical, technological, social and cultural contexts of learning. While we found the design principles enumerated above useful in developing the online KBAI class, we expect that their operationalization will vary across different learning contexts.


A. Goel. Preliminary Evidence for the Benefits of Online and Blended Learning. In A. Madden, L. Margulieux, R, Kadel & A, Goel (editors), Blended Learning in Practice: A Guide for Practitioners and Researchers, MIT Press. April 2019.

A. Goel & D. Joyner. An Experiment in Teaching Artificial Intelligence Online. International Journal for Scholarship of Technology-Enhanced Learning, 1(1): 1-27, 2016.

A. Goel & D. Joyner. Using AI to Teach AI: Lessons from an Online AI Class. AI Magazine, 38(2): 48-59, Summer 2017.

A. Goel & L. Polepeddi. Jill Watson, A Virtual Teaching Assistant for Online Education. In C. Dede, J. Richards & B. Saxberg (editors), Education at Scale: Engineering Online Teaching and Learning, NY: Routledge, 2018.

C. Ou, D. Joyner, & A. Goel. Designing and Developing Video Lessons for Online Learning: A Seven-Principle Model. Online Learning Journal, 23(2): 82-104, June 2019.

Author Biography: Ashok Goel is a Professor of Human-Centered computing in the School of Interactive Computing at Georgia Institute of Technology. He is also the Chief Scientist with Georgia Tech’s Center for 21st Century Universities. In 2014, he co-developed an online course on Knowledge-Based AI; in 2016, his research laboratory developed Jill Watson, a virtual teaching assistant for answering questions in online class discussion forums; and in 2019, he co-edited a volume on Blended Learning published by MIT Press. In 2017, Ashok received Georgia Tech’s Class of 1934 Outstanding Innovative Use of Educational Technology Award; in 2019, he received AAAI’s Outstanding AI Educator Award; and in 2020, he received the University System of Georgia’s Faculty Hall of Fame Award for Scholarship of Teaching and Learning.