Machine Learning and AI have a lot of success players and tangible examples. Just think about the self-driving cars or self-flying drones. According to Nils Jansen, many people believe that AI can solve all problems. However, there are still huge problems that remain. ‘People are still better than machines in solving a lot of problems. The question is: are AI and machine learning doing what they intend to do? In other words: do they solve the desired problems? Our research focusses on those questions. We can, for example, predict the weather, based on certain data. What we also know, is that there are things that can change the outcome. We intend to become more robust against all possible uncertainties that occur in the real world.’
Reinforcement Learning
Reinforcement Learning is in fact a very particular machine learning technique. What it does mainly, is exploring the environment. Nils Jansen clarifies: ‘The problem is that the agent, for example a self-flying drone or self-driving car, often tries out every possible action and then observes what the possible outcome is. The most famous example is Alpha Go. A couple of years ago, the AI computer that played the game Go beat the best human player in the world. It can thus be very powerful, however: it is also extremely safety critical. It can harm itself and its environment, because it tries all possible actions.’ Thus, contrary to what most people think, the new technologies are not always better and smarter, but also create new safety challenges.
Making it safe again
The European Research Council is part of the EU and offers big research funding programs. Nils applied for a starting grant beginning of this year, so he could do more research into the safety problems of Reinforcement Learning. And he succeeded: ‘I was granted € 1.5 Mio to invest in making Reinforcement Learning and AI safer. We use mathematical models to precisely capture the uncertainty that happens in the real world. Our group tries to measure what will happen in the worst case, but then more reliable. For instance: we assume the wind will have a certain speed and direction, this is based on historical data. What we add to this, is that we precompute the behavior which is safe, based on the incomplete data we have. Amazon will use drones to deliver packages in a couple of years. At this moment a drone has data which it can use to prevent crashing into a building, like data about the wind. However: those are not reliable enough. Therefore we precompute the worst thing that could happen. After that, we use this information to make a decision with an acceptable amount of risk.’
The funding helps Nils Jansen and his team working on the reliability of AI and machine learning. And eventually, that makes the world a safer place for all of us.