It is a fact that Google is at the forefront of Artificial Intelligence research, and when an IIIA alumnus is part of its team, it is a source of pride for us. Ewa Andrejczuk (Poland, 1987) studied Computer Science at the Warsaw University of Life Sciences and finished her industrial PhD in 2018 and was part of the starting team of EduTeams, the application to compose synergistic teams. She visited us a few days ago, and we were lucky enough to interview her.

Let's start; how did you begin your story with Artificial Intelligence?

It's an interesting story. I finished my master in Poland, and I had not much to do with AI. During my studies, I had only one class, Machine Learning. I liked it, but I didn't think about it, and at the time, I was working in the data engineering field. I was working on ETL processes when I started getting into AI. I was always interested in understanding how things work. In particular, at that time, I wanted to know how Google browser worked. I began to read a little about this, and I went through the (now most popular) AI course in Coursera with Andrew Ng called Machine Learning. I was super excited, and I said: "Okay, I really want to work on that." So, I got the idea to start the PhD, and yeah, I was looking for a suitable program. I sent my curriculum to IIIA, and then Carles contacted me and asked if I was interested in that opportunity and hell yeah! (laughs). So, I landed in IIIA. What I did with Carles Sierra and Jar (Juan Antonio Rodríguez) was the industrial PhD. It was a good program for me because I already knew how the industry worked.

So, you did an industrial PhD; could you explain the difference with an ordinary PhD?

There is a program called Doctorats Industrials sponsored by the Catalan Government. The difference between an ordinary PhD program and the industrial one is that the company hired me as a consultant. I had to attend fewer classes, only 60 hours in total. Every week, I spent three days at the institute with Carles and Jar and two days in the company working from there. My focus was my research, and I wasn’t obliged to do anything different for the company. This program was an excellent way to bridge the gap between industry and academia.

What was the topic of your PhD?

I did my PhD mostly about team composition in Multiagents Systems. So, what we tried to do currently is called Eduteams. I started that line of research. In the company where I worked, they already had some trust algorithms and wanted something similar for the teams. They wanted to know how to compose effective teams and understand how good each team was. So, we worked on that idea with Carles and Juan Antonio. As the first step,  we designed a function that measures how good (aka ‘congenial’) a team is based on the personalities and genders of team members. Second, we developed algorithms to build congenial teams. Then we realised we didn’t want to have only one strong team and other weaker teams. We wanted all teams to perform well. So, we developed algorithms to partition a set of individuals into congenial teams. That is, balanced in personality and gender. Finally, we wanted team members to have the required competencies to perform a task, so we added that notion to the model. We called the teams composed with this model synergistic. This is how we came up with these algorithms and the solution (EduTeams).

How was the process?

The company wanted my work to be for humans rather than agents, so I started reading about psychological literature and seen what is essential in team composition and team performance. We formalised a psychological theory mathematically. That is, we built a utility function based on the findings of Stanford professor Douglass J. Wilde. His idea was only about personalities and genders. Besides, many sources were also talking about competencies, so we decided to add them as well. So, it was like this combined part with personality and gender and then was this prophecy part with competencies, and we stick it together. That was our utility function. I thought in the beginning that the problem would be easier. With time I learned that having all teams balanced in this utility function that I built: personality, gender, and competencies was actually NP-complete.

How was your experience in the IIIA?

It was the best experience ever. There is no other place like that. After I had been there, I went for a post-doc and talked to many labs, many colleagues, and I can say IIIA is unique. Now I am at Google. Maybe Google is a bit similar in that everyone is amicable and helpful, and there is no hierarchy. So, you can approach anyone, even someone much more experienced than you, and they will discuss with you and try to help you. IIIA is a bit like a family. I enjoyed working with Carles and Jar, and every time I had a problem, they always had time for me. If I didn't know how to solve something and I was stuck, I dropped by Carles office, and he helped me by asking questions or looking for solutions with me…

Now you are working by Google, tell us a bit of your work.

I am working on reinforcement learning. I am in a program called "AI Residency". It's a one and a half year program to ramp up on machine learning research. I was very excited to get to the program. During my time at Google, I can select my projects and work on things that I am excited about.

In which consist of reinforcement learning?

Typically, in reinforcement learning, we have an environment and an agent that interacts with it. A typical example of a domain is Robosuite , where an agent performs manual tasks such as pick up a can and place it in a dedicated space.

Usually, at the start, the agent doesn't know anything about the environment. Only by interacting with the environment the agent can learn and understand what the rules are. For instance, in robosuite, the agent needs to learn how to behave to maximise a reward. It gets a reward after every action it performs. For example, if the agent moves the arm and gets closer to a can, the reward could start growing, and the agent understands that this is a good action. There are many exciting questions in reinforcement learning that are an active area of research. The primary purpose is to design an agent that learns efficiently, similarly to humans. Agents can also learn by observing and trying to imitate experts such as humans or other agents. 

To conclude, what would you say to someone who is starting in Artificial Intelligence?

Go for a PhD to IIIA (laughs). I would say they should be curious. Read to get solid foundations, go to conferences and discuss with people, and code a lot! And I would advise to do a PhD cause it opens a lot of doors. Without a PhD in AI, it isn't easy to be an AI expert, it may limit the career options. So go for a PhD; it would be my advice.