Bayesian optimization

When designing new products there are often sequential tests of different designs in order to find the best one. These individual trials can be expensive to carry out. For example, when designing an aerodynamic part for a car, each new design may require a new, single use, prototype to be built and tested in a wind tunnel. Reducing the number of candidates tested before finding the best one would therefore be very beneficial and one possible way of doing this would be to use machine learning, and in particular Bayesian optimization. This would tie in with EPSRC's engineering theme.
We therefore wish to explore new techniques or applications for Bayesian optimization. There are many interesting avenues of research we may like to look at such as:
- new Bayesian optimization algorithms;
- methods for the creation of complex objective functions (trading off different objectives);
- exploring using new generative models which are able to propose new designs rather than just assessing pre-defined candidates.
One particular application we may like to explore would be to look at using Bayesian optimization in the process of drug discovery, tying into EPSRC's healthcare technologies theme. The development of an approved prescription medicine is estimated to take over a decade and to cost $2.6 billion according to the Tufts CSDD. Part of this cost comes from the many compounds that are tested in the development of a new drug, as the ABPI (Association of the British Pharmaceutical Industry) says around 25 000 compounds are tested, but only around 25 enter clinical trials, for each successful drug found. One way we could try cutting down this number by using Bayesian optimization would be to model the posterior of the objective function (for example the posterior of the drug's effectiveness) over the candidate molecules using Gaussian processes and graph kernels. Alternatively or in addition, we may like to explore and extend other new methods researchers have been using for getting features from graphs, such as using neural networks.
This research would mainly fall under EPSRC's artificial intelligence technologies research area, as we wish to develop new machine learning techniques and apply them to the process of Bayesian optimisation. However, it is also related to:
- Engineering design, as are looking at new ways to go about testing possible designs to find the best one at a certain task,
- Statistics and applied probability, as we will develop and use probabilistic models for modelling the posterior of the objective function.
- And also potentially computational and theoretical chemistry, as we could look into applications applying computational and mathematical methods to relate chemical molecules' properties to their structure.