Automatic Experimental Design with Human in the Loop

This project is related to Turing AI World-Leading Researcher Fellowship: Human-AI Research Teams - Steering AI in Experimental Design and Decision-Making
Machine learning offers great promise in helping us solving problems by automatically learning solutions from data, without us having to specify all details of the solution as in early computational approaches. However will still need to tell machine learning system what problems we want them to solve, and this is currently undertaken by specifying desired outcomes and designing objective functions and rewards. From formulating the rewards for a new problem is not easy for us as humans, and is particularly difficult when we only partially know the goal. As it's the case at the beginning of scientific research. In this programme we developed ways for machine learning systems to help humans to steal them in the process of collecting more information by designing experiments, interpreting what the results mean, and deciding what to measure next, to finally reach the conclusion and Chester with the solution to the problem.
The machine learning techniques will be developed first for three particular important problems and then generalised to be broadly applicable. First is to diagnosis and treatment decision making in personalised the machine, and the second steering of the scientifically experiments in the synthetic biology and drug design, and the third design and the use of digital twins in designing physical systems and processes. And an AI centre of excellence will be established at the University of Manchester, in collaboration with the Turing institute and number of partners from the industry and healthcare sector, and with strong Connexions to the networks best national and international AI researchers.

A brief summary of the project is:
We are working to design an AI assistant that can help other agents solve experimental design problems.
The design problems we wish to solve can be considered as sequential decision making problems - at each iteration, the design is adapted slightly.
Real world experimentation is infeasible due to factors including monetary cost, ethics and time. We must therefore simulate experiments in a digital twin.
At each iteration, the AI assistant provides a recommendation to the agent, which can either be accepted or rejected.
The agent's utility function (goals) are unknown to the AI assistant apart from some limited prior knowledge. These goals must be inferred from accepting/rejecting AI recommendations.
Using a city planning/traffic reduction scenario for the application of my project. We make changes to the network to reduce traffic and air pollution.

The goal is to generalize Bayesian automatic experimental design to multi-agent models consisting of an AI assistant and the human user, resulting in the AI assistant being able to decide its next actions. Tentative solutions involve developing fast probabilistic surrogates for existing simulator-type models and experimental design with approximate inference. The student will work alongside a team of researchers, supervised by a machine learning expert, and will have access to exciting application opportunities in both companies and academia.