Information networks and ecological networks and learning control policies in probabilistic Boolean networks seminar

The 'internet of things', Edge and Cloud computing all have important contributions to make to a wide range of application areas. We are seeing much talk about their application to smart cities, healthcare and Industry IV, for example. However, there is one area where they have potential to help catalyse a significant, and much needed transformation. That is agriculture.

We are, of course, seeing projects on precision agriculture with useful reductions in seed, fertiliser, pesticide and herbicide wastage. But we can do better than that; much, much better. A body of theoretical and empirical studies around the world is now showing that if agricultural practitioners work with natural processes to build both above and below ground biodiversity, then this can enhance the provision of ecosystem services that help to capture soil organic carbon, recycle trace elements and the core NPK needs, encourage predators of weed seeds, and more.

The term ecological engineering is now becoming current to encapsulate a methodology by which agriculture could be transformed away from a net contributor to greenhouse gas emissions, groundwater pollution and disruption of social cohesion in rural communities. Instead, it could become a major vehicle for carbon capture and storage, and provision of clean and safe drinking water, whilst maintaining yields, enhancing the quality of food and increasing employment opportunities in rural areas.

Facilitating this will need important contributions from the Computer Science community. This talk will provide an overview of what can be done now, and what is still needed in order to provide real time monitoring of agroecosystems to support the transition of global agriculture into a scientifically managed and sustainable system.

Boolean networks (BNs) and probabilistic Boolean networks (PBNs) were introduced as a computational model for the study of gene regulatory networks (GRNs) and, more generally, of discrete dynamical systems. The ability to control such networks is important, e.g., in targeted therapeutics. Control here refers to the process of making strategic interventions to the state of a node (gene) in order to direct the whole network to a state that exhibits favourable biological properties.

In this talk, I will draw from reinforcement learning, a particular branch of machine learning that focuses on sequential decision making under uncertainty, to develop control policies that drive a PBN towards a target state, typically an attractor, within a finite number of steps. The control method is model-free, hence does not require knowledge of the network dynamics, and is therefore suitable for applications where inference of such dynamics is intractable. I will present some experiment results on synthetic PBNs but also PBNs constructed directly from gene expression data (metastatic melanoma).

Finally, time permitting, I will also touch upon inferring the network in the first place.