Machine Learning Evaluating Evolutionary Fitness Depending on Initial Conditions
Oleg Kuzenkov, E. Ryabova, A. Garcia · 2021
The purpose of the work is to evaluate evolutionarily fitness depending on the initial state of a biological population based on machine-learned ranking. We apply a novel computational approach using artificial neural network technologies to reconstruct the fitness function in both theoretical models of population dynamics and empirical biological systems from data. Our approach uses a long-time population series (obtained either from the model or from data) and establishes the ranking order of inherited strategies which reflect their selective advantages for fixed initial conditions. We apply our method to explore the evolutionary stable diel vertical migration (DVM) of zooplankton, the phenomenon that is considered as the most significant synchronous biomass movement on Earth. We use the classical predator-prey model with a logistic prey growth. We approximate fitness as the linear function of a few key parameters by solving the classification problem. To do this, we create learning and testing samples. We find the evolutionarily stable (optimal) strategy by maximizing evolutionary fitness for different initial conditions using the obtained approximation of the fitness function.