Over the last few decades developments in AI have taken it from playing chess to becoming an indispensable part of rapid drug development pipelines. The effect on industry and society has been transformative. However the potential of AI is currently constrained to a minority of problems, those where we can precisely specify an objective or have plenty of good solutions that can be learnt from. In this project we focus on other problems, where no good descriptions of objectives or optimal behaviour are available. This means that we must keep human judgment in the loop, to give feedback to the AI and guide it, while keeping the AI as autonomous as possible. To do this we create AI that automates what it can and relies on the human only when it is uncertain about how it should behave.
For such a human-AI team to be efficient, the AI must be aware of what it does not know about the goal, so that it can figure it out by asking for information from the user. Thus, the AI needs to know what and how to ask the user to minimize the user’s effort through automation. In this project, we will introduce AI-assistants with advanced user models. With such models, the AI-assistants can design effective and efficient interaction with the user to elicit as much information as possible about the goal, and thereby acquire long-term decision-making skills through reinforcement learning to automate the solutions to the extent the goal is clear.
We will deploy these AI-assistants on real-world reinforcement learning problems where designing the reward functions is difficult. In particular, we will work in collaboration with companies working on chemical design and autonomous driving, and use their real-world pipelines as our test benches.
At present, road transport contributes a significant amount to the total carbon dioxide (CO2) emissions in the EU. Thus, cities look for practical strategies to make their transport system more efficient and sustainable. Electrification of road transport is the primary technological change needed to meet the carbon reduction targets. However, electrification is unlikely to be sufficient since the electricity production will not be carbon neutral in the near future. There is a second major technological transformation on-going in road transport—digitalization—bringing forth the advent of connected automated vehicles. Connected automated driving will transform traffic flow management into a proactive disaggregated and cooperative paradigm that, via appropriate strategies, may enable a decrease in CO2 emissions. However, the total joint effects of electrification and autonomy on CO2 emissions are not well understood. There are major potential cross-effects, such as an increase of vehicle-km travelled due, for example, to an autonomous car visiting a recharging station. Furthermore, the transitions will not be instantaneous but electric and combustion engines, and automated and human-operated vehicles will co-exist for a significant period, which is not typically taken into account in existing studies. When combined with new possible vehicle ownership models and policies, the impact of automation on urban traffic remains highly uncertain. In the future, digitalisation and communication technologies may also enable much more flexible management of the existing infrastructure.
AI4LessAuto brings together atmospheric and computer scientists and traffic engineers in active dialogue with municipal stakeholders with the ultimate aim to understand how autonomous electrified traffic should be organized during the transition period in order to reduce CO2 emissions. This is achieved by 1) building a framework of computational modelling tools to evaluate the CO2 emissions originating from electrified automated vehicles, and 2) developing artificial intelligence based control from vehicle-level to city-center wide traffic-level in which CO2 emissions are minimized.
We address the problem of autonomously learning to perform tasks. This is the fundamental challenge in developing genuinely autonomous systems across fields, from robotics and autonomous vehicles to customer services, employable outside the narrowly defined tasks the current hand-crafted solutions require. The challenging bottleneck is that autonomous learning necessarily involves exploration, which in the real world is costly and unsafe. This is the reason why the impressive demonstrations of recent deep learning powered reinforcement learning are not yet available in the real world. B-REAL aims to bridge this reality gap between autonomous learning in real and simulated worlds.
The core idea is to learn an artificial surrogate model, in which both performance and safety can be checked before real-world action. What is new in this project is that now we believe we are in the position of setting up the currently disjoint modelling methods required for coping with the reality gap into a complete next-generation autonomous system. This requires efficient ways of learning accurate simulators, matching them with the real world, and including safety constraints. Safe and effective operation requires taking into account the uncertainty of the simulation outcomes, which is now possible even for the necessarily flexible simulators due to our recent, award-winning breakthroughs in deep probabilistic modelling.