Coherent inference over a network of probabilistic systems for decision support with applications to food security.

Different panels of experts each provide their judgments together with supporting evidence through the outputs of their own probabilistic models. These then need to be networked together via their inputs and outputs to provide a unifying inferential foundation to underpin decision support. The different panels of experts share as common knowledge qualitative features of the domain often expressed as an agreed graphical model. A subjective expected utility maximising policy maker now plans to adopt the probabilistic judgments of the appropriate domain expert panels as her own. She will then use this unifying statistical model to draw together the disparate evidence into a single coherent and defensible whole and use this to guide her policy evaluation. We will develop this formal and defensible statistical methodology. Thus we will find sufficient conditions for different types of dynamic graphical models to remain coherent before and after relevant evidence is accommodated into the system. When the integrity of the system is violated we will develop methods which appropriately measure the extent of this violation. Recent advances in algebraic statistics will be used not only to investigate how such demands for coherence might be relaxed when the algebraic form of a user's utility function is known but will also guide associated computational algorithms. The development will be directed by the needs of policy makers and the structure and information bases used by experts and researchers into UK food security

This programme will develop some of the technologies and methodologies that will make it possible and viable to design customized decision support tools. This will be able to demonstrably and quickly provide integrated probabilistic models for various social policy analyses and guide policy makers with evidence based support to ameliorate the threats it faces . The PI has been involved for 15 years in developing probabilistic decision support tools to manage nuclear risks. The PI and the VRF Prof. Nicholson also share a long history developing Bayesian Networks to work in dynamically evolving environments for clients in industry and government (see c.v.'s). This project will directly address a new area needing proper integrated decision support: food security and indirectly decision support methodology within analogous domains. By contributing to better governmental and international decision making it will contribute to the health, safety and security of people in the UK and worldwide. The theoretical researchers will work with two CI's who have excellent domain knowledge of complementary areas of food security research - one of the foci of the study and have excellent links to potential users of the new technologies we plan to address in this programme.

We envisage an impact of this research on:
(i) Applied Statisticians needing to inform decision support systems for policy analysis over networks by using the developed methodologies to do this more effectively. This impact will be felt in the short term (0-5 years).

(ii) Decision and risk analysts. Methods for appropriately integrating diverse probabilistic systems, with online updating of information bases supported by appropriate inferential methodologies will continue to be developed and implemented in a wide range of applications. This research will make a significant contribution to this field in the short and medium term (0-10 years).

(iii) Food researchers. The two studies we will undertake in this programme will form a template for the subsequent more comprehensive analyses designed to integrate decision making over the whole system. We would expect this impact to take place mainly in the medium term (5-10 years).

(iv) Policy makers and food providers. Our methods will eventually enhance understanding of the threat of food shortages and the development of short and medium term strategies enabling more efficacious and comprehensive policy decision making. This could have very strong societal impact and potentially benefit the lives of many. We would hope this will be felt in the medium term (5-10 years).

The types of impact that we expect from this project are (using the EPSRC impact categories):
1 Impact on Knowledge, in particular on techniques (through beneficiaries in groups [i] and [ii])

2 Impact on Society, in particular impact on the general development of better networked decision support systems through beneficiaries in group [ii] above) and on the associated food researchers [iii] and policy makers (group [iv] above, but also others in analogous roles in different domains.

3 Impact on People, in particular on skills through training the RA involved and facilitating the involvement of postdoctoral researchers in the workshops.