Apply now! 3-year Postdoc on building trusted AI for environmental decisions
We are seeking a highly motivated and dynamic postdoctoral research fellow to work on developing AI specifically designed to deliver interpretable and explainable solutions to environmental decision problems.
The PDF will be supervised by iadine Chades (CSIRO, Brisbane Australia, Google Scholar), Tom Dietterich (OSU, Corvallis, USA, Google Scholar) and Andy Reeson (CSIRO, Canberra, Australia, Google Scholar).
Salary range: AU $82k to AU $93k plus up to 15.4% superannuation
Location: CSIRO Dutton Park, Brisbane (EcoSciences Precinct).
Apply now! Application closes 31st of October.
Environmental managers seldom have the luxury of full information to guide their decision-making. In conservation and fisheries management, species are often too difficult to detect to provide accurate population abundance estimates needed to inform decision-making. Managers must make decisions despite the uncertain outcomes of their actions, or risk failing to achieve their goals through inaction. Making decisions under uncertainty is a complex mathematical problem that can be efficiently solved using Artificial Intelligence. For example, where the future is uncertain, managers must adapt their decisions as they act, using feedback from their observations to predict optimal future actions while reducing uncertainty over time. This adaptive management, or ‘learning by doing’ can be optimised using powerful AI decision models such as Markov Decision Processes (MDP) and Partially Observable Markov Decision Processes (POMDP). However, optimisation alone is not sufficient for good management.
AI decision models will only be useful if they are used by decision-makers. To date, most attention has been placed on the technical aspects of AI, with little emphasis on their adoption by human managers. This is a widespread problem, but is particularly acute in the environmental domain, in which decision-makers are typically trained in biology or environmental sciences, and experienced in practical fieldwork rather than technology. This has resulted in environmental managers not taking advantage of the opportunities these decision tools could offer to tackle complex environmental decision-making.
Over the last ten years, our research has pioneered the use of Artificial Intelligence decision tools to manage our environment and our science has been recognised by prestigious publications in both ecology and AI conferences (AAAI, IJCAI). Greater impact will come when more and more environmental managers effectively exploit the benefits of AI. We have identified that our greatest need is to develop trusted easy-to-use Artificial Intelligence for environmental managers rather than Artificial Intelligence that focuses on optimal solutions. Such solutions must be easy to interpret and provide explainable mechanistic insights. They should also help users to learn and explore the impact of alternative management actions and scenarios in order to add value to their judgement.
This requires an understanding of human decision-making, including biases and heuristics that could result in AI outputs being misinterpreted (or ignored) by environmental managers. The field of behavioural economics has identified characteristics of human judgement and decision-making, as well as elicitation techniques to minimise bias, which are currently poorly accounted for in AI decision tools. Building on behavioural economics, the PDF will address a key question in environmental management: How should AI be designed and implemented in order to be considered a trusted advisor by managers? What information should AI elicit, and how should it present and explain its outputs? Developing AI algorithms that can provide such solutions in human-operated systems is a substantial task that will be the focus of the PDF over the next three years.