Linking non-parametric upscaling of agent-based models to machine learning

WIEGAND, T.; CIPRIOTTI, P.A.

Leipzig

Conferencia; European Conference on Ecological Modelling; 2023

ISEM / UFZ

The problem of scaling up from tractable, small-scale observations and experiments to prediction of large-scale patterns is at the core of ecological theory and application, and one of the central problems in ecology. We present a general non-parametric framework to upscale spatially-explicit and agent-based models to predict landscape scale dynamics based on individual-level information, but without simulating the computational demanding interactions among individuals with brute force. The idea is to design a state space, defined by the important state variables of a small-scale model describing the dynamics within a patch of hundreds or thousands of individuals, and to divide it into a finite number of discrete states. Similarly, the external drivers are divided into a finite number of discrete conditions. For a given time interval, transition probabilities between states are then tallied by monitoring extensive simulation runs of the small-scale model that need to cover the entire range of initial conditions, states and external drivers that may occur for the desired application. If the space of external drivers is not too high, the original small-scale model can be replaced by the transition probability matrices, resulting in Markov chain models. However, in the more general case of a higher dimensional space of external drivers, machine learning, such as deep neural networks, are a promising approach to estimate the transition probabilities in dependence on the environmental variables. A hierarchical perspective then embeds the transition matrices derived from the small-scale model into a larger scale model where the dynamics of the states of each patch are driven by the transition probabilities, depending on the environmental variables, and additional larger-scale rules govern possible spatial interactions among patches. This general upscaling approach uses detailed agent-based model simulations to train a meta-model based on states and transition probabilities and therefore accounts for the full complexity of the agent-based model, but views the system only through the filter of the states that reflect what is regarded as important at the largest scale.