Iridescent founder and CEO Tara Chklovski recently sat down with Cristina Conati, a professor in the Department of Computer Science at the University of British Columbia. Professor Conati is interested in creating intelligent interactive systems that can adapt to individual users’ needs. One application of this is educational games that need to both entertain and teach their players.

Cristina ConatiTara Chklovski: Thank you for sitting down with me! Tell me about the problem or area of research you are working on.

Cristina Conati: In general I’m interested in the idea of user-adaptive interaction, which is using AI techniques to create tools that can personalize their interactions with users by capturing the user’s needs and preferences as they interact. Within that, I’ve done work on intelligent educational games, which are computer games that have a pedagogical purpose. There are activities that are designed to engage and amuse the players, but also to teach specific concepts – for instance, mathematical concepts.

AI technology in educational games enables intelligent, ongoing personalization, and the ability to gauge whether the player is engaged and whether they’re actually learning. That’s important because there are a number of educational games that are engaging but are not always effective at teaching. Creating engaging activities that are also effective at teaching is difficult through design alone, because players are different – what’s engaging for one person is different for another. The same is true of teaching strategies. But if we can create educational games that can adjust in real time to suit the person playing because it can understand their abilities, personality, and preferences, then students will have better learning interactions with the games.

Tara Chklovski: What sort of data do you collect to evaluate the different states of the user?

Cristina Conati: For one example, I worked on an educational math for number factorization where players climb mountains by moving to different zones marked by numbers. Pairs of players collaborate in climbing together, and must move to numbers that do not share common factors. To do this, they have to understand concepts of number factorization. Based on the moves a player is making, the game assesses which relevant math concepts the players understand and where to provide pedagogical support. The game also assesses emotions the players are feeling at any given point – is the player experiencing positive emotions towards the game? Or is she frustrated with the game? – because such emotions are important to provide effective pedagogical support.

This support is provided by an AI agent portrayed as a funny little character that gives help and hints on how to play the game, and how to understand number factorization. The agent tries to understand which kind of affective reactions the player is having towards the help and suggestions that it provides. We’re still working on the best way to make the agent change its behaviors to counteract the student’s negative emotions, such as frustration, or to get the student to make moves that are more related to the mathematical concepts the game is trying to teach.

Tara Chklovski: What would the child do in the game that would signal frustration?

Cristina Conati: Well, with educational games, one problem is that it’s possible to earn points in the game without actually thinking about the concepts the game is trying to teach. So in this game, it’s possible for a player to figure out how to move up the mountain without understanding the mathematical concepts. But the little agent has ways, based on AI techniques, to guess that this is the case and so it might say “You made a correct move, but are you sure you understand how to factorize this number?” That comment could be frustrating for the player, because they could say, “I made a correct move; I just want to play the game and earn points and get to the top of the mountain. I don’t care about how to factorize this number.” But based on its evaluation of the player’s behavior, the agent might decide that it still needs to make this comment to trigger student reflection on the mathematical aspects of the game, even though it could cause frustration.

Essentially, the AI model guiding the agent’s actions uses data from a variety of sources, including sensors that detect affective signals from the moves the player makes, and it uses those data sources to make a prediction on how well the player is learning and on their emotions. It makes uses of AI techniques that allow the system to make inferences when there is uncertain information. Uncertain information exists all the time, even with human-to-human communication. For instance, I’m talking to you now and I don’t have a lot of information about whether you’re understanding what I’m saying. You might ask me to clarify, but at the moment, because I can’t see your face, I don’t know if you understand. So as I’m talking I’m making a model of what you might be understanding and trying to adjust what I’m saying based on that model, but it’s not certain, since I don’t know you that well and can’t see you. It’s exactly the same thing for an AI agent – it’s working in uncertain conditions.

Tara Chklovski: Switching topics a bit, why did you get into this field?

Cristina Conati: That’s a good question! I’m very interested in how AI can have a real impact in our daily lives, and education is a very important element of our life and society. Devising intelligent agents that can provide some of the benefits of one-to-one instruction is beneficial. From an AI point of view it’s very challenging because if you think about it, in order to be able to be a good teacher you need to be able to have a lot of knowledge. Not just about the subject area that you’re trying to teach, but also how to capture important signals from the learners about whether they understand, whether they’re engaged, and how they’re reacting to you and the material. It’s challenging because there is a lot of uncertainty.

There is a lot of data we can collect, but understanding how to interpret this data and how to reason out the good decisions that you can make is very difficult from a technical point of view. But I’m passionate about AI and passionate about the challenging applications of it. And if it succeeds it will have an impact.

Tara Chklovski: What are some challenges you face, and how do you overcome them?

Cristina Conati: Uncertainty is one challenge, and another problem is that in order to develop and deploy these technologies it’s important that the schools are willing to receive them. And it’s very important that the teachers and the educators are involved in the design process. I might be an expert in AI and able to generate models that can work well in uncertain conditions, but I need teachers to share their thoughts on how the system should be designed so that it fits well in a specific classroom or school or curriculum. Unfortunately, often there are no resources to do this type of collaborative work, because teachers are busy and they have many other things that they have to do. And even if educators are willing to introduce new learning technologies in schools, sometimes in practice it’s difficult because it’s hard for the educators to find the time to be involved as extensively as they want to be and as they should be. It’s definitely not their fault! It’s not a criticism at all, it’s just a fact.

Tara Chklovski: It’s been so interesting to see how quickly tools are being developed with those sorts of concerns in mind and where they are and aren’t succeeding. Thank you so much for your time Cristina, it’s been great to talk to you!

This interview is part of ourAI in Your Community series.