Intelligent Earth Seminar: Richard Bailey (SoGE), University of Oxford

12 November 14:00
Seminar Room 1a, Doctoral Training Centre, 1-4 Keble Rd & Online

 

richard bailey

Richard Bailey is Professor of Environmental Systems in the School of Geography and the Environment, and Fellow of St. Catherine’s College, Oxford. His research interests are in the dynamics of natural environmental systems and human-environment interactions over a range of timescales and contexts. A broad range of research methods are used in this work, including numerical and analytical modelling. Much of Richard’s work is influenced by Complex Systems research, particularly in regard to modelling human-environmental systems (oceans, plastic pollution, and others). Current interests include resilience in complex socio-ecological systems, agent-based modelling, AI and machine learning theory and applications.

Simulating intelligence in environmental systems: WHAT, HOW and WHY.

Environmental systems contain myriad intelligent processes spanning physical, biological, ecological, and social settings, across multiple organizational and temporal scales. These processes lead to rich, hard-to-predict and hard-to-understand behaviour. Including some representation of this intelligence within our models of environmental systems is therefore essential. 

WHAT counts as intelligent behaviour in human-environmental systems is up for debate, and the lens for this seminar will be the information-processing capabilities of processes associated with emergence. To provide structure to these different levels of intelligence, I propose the following framework: Tier 1 intelligence is reactive, ‘pre-compiled’ control (e.g. evolved priors and rule-based responses); Tier 2 is adaptive, typically model‑free learning (e.g. foraging behaviour with working memory); Tier 3 is deliberative, model‑based planning with memory (e.g. strategic reasoning and counterfactual evaluation). 

HOW this plays out is illustrated using examples from previous and ongoing work involving emergence in cellular automata, continuous learning in agent-based models, and self-editing LLM-agents that use ‘conversation training’ to improve reasoning over non-verifiable “wicked” environmental problems.

WHY: embedding intelligence into models of environmental systems provides richer representations of their adaptive dynamics and facilitates production of more reliable ‘flight simulators’ for managers and policy-makers. Examples of concrete applications will be discussed.