top of page

Machine Learning, Adaptation, and VRCs

Updated: Feb 6, 2020

In many ways, the Higher Ground Foundation can be said to be a part of the “risk community”, which is exemplified by traditional insurance companies but, in recent years, has expanded to multidisciplinary groups interested in decision-making under uncertainty, including the Society for Decision Making Under Deep Uncertainty and the UK’s Analysis Under Uncertainty for Decision-Makers (AU4DM) Network. This week, the AU4DM held a session in London on “Opportunities and limitations of machine learning and other AI techniques in complex decision making” that several HGF principals attended.

Machine learning, a currently hot sub-field of research into what is more popularly referred to as “artificial intelligence,” is being applied in numerous and often surprising corners of life, including the personalization of video recommendations by streaming services and the growing use of sophisticated image and reverse-image searching applications. It is also being increasingly used as a tool to address big “risky” issues with inherent uncertainties and fuzzily defined outcomes.

So why is HGF interested in decision-making tools like AI? In a number of ways, the climate adaptation challenge can be appreciated as a complex problem space requiring the simultaneous optimization of multiple solutions, many of which are specific to individual project types. In this sense, it differs from conventional climate mitigation approaches, which, though in themselves challenging and necessary, share the single optimization dimension of minimizing greenhouse gas emissions (conversely, maximizing capture and storage).

Adaptation frameworks such as the VRC Standard tackle a more complex set of problems. Projects and methodologies are defined along multiple dimensions that are each populated by different types of input (e.g., climate factors; project geographic, social, and economic parameters; planned intervention technologies and techniques) that must be extracted from big data sets and processed together to produce a single optimized output: the number of VRCs generated by a project.

In some classes of project, this can be done straightforwardly through the plugging of easily accessed, pre-formatted data into standard templates. Other adaptation projects are more complex and not as tractable to a standardized methodological approach. Take the problem of river valley or coastal flooding as an example: each vulnerable community or locale will face a unique threat driven by local climate factors, property/infrastructure distribution, and specific basin/coastline topography and hydrology. To provide comprehensive VRC coverage across a region or nation using a standard project design approach would involve the aggregation of countless unique manual solutions, which becomes very difficult.

In this respect, ML approaches have the advantage of being generalizable through training based on large sets of unlabelled data. In the flood defence case, a neural network could be trained to recognize the individual parameters of a general flood-prone area based on learning from visually coded map data and then apply the learned rules to the appropriate modelling and optimization of individual project locales. Once the relatively costly job of training the model has been done, the VRC project definition process can be replicated by the network at scale without requiring direct expert input (although reality-checking will of course be necessary).

Although academic researchers have already begun to take an interest in the application of machine learning to climate issues, much of the focus has been on more limited cases, and significant work would be needed to synthesize the work done so far before the first “VRC deep learning network” could go online. This will constitute the HGF’s own (supervised and unsupervised) learning process, and we will continue to closely follow the field and look to organizations such as the AU4DM and others for meaningful partnerships and insight.

60 views0 comments


bottom of page