Genetic AI: Evolutionary Games for ab initio dynamic Multi-Objective Optimization
Philipp Wissgott
TL;DR
Genetic AI introduces a parameter-free, ab initio framework for multi-objective optimization that converts input data into a fixed population of genes and organisms and evolves their fitness via replicator dynamics governed by four evolutionary strategies: Dominant, Altruistic, Balanced, and Selfish. The method updates gene fitness $\bm{\gamma}$ and organism fitness $r_i$ without training data, using a linear organism fitness $r_i^{(k)} = \bm{\omega}_i \cdot \bm{\gamma}^{(k)}$ and strategy-specific local updates $\Delta^{g}$ and $\Delta^{\omega}$, with optional self-consistent mixing of strategies. It provides ab initio, predefined, or training-based mixing of strategies to yield nontrivial evolutionary equilibria and demonstrates its mechanics on simple and real-world flight-selection problems, revealing how data features become more or less relevant under different dynamics. The work argues that Genetic AI can reveal correlations, symmetries, and universality in data, offering a new lens for optimization and data analysis with potential extensions to search, recommendation, prediction, and hybrid systems with neural networks or large language models.
Abstract
We introduce Genetic AI, a novel method for multi-objective optimization without external parameters or predefined weights. The method can be applied to all problems that can be formulated in matrix form and allows for a data-less training of AI models. Without employing predefined rules or training data, Genetic AI first converts the input data into genes and organisms. In a simulation from first principles, these genes and organisms compete for fitness, where their behavior is governed by universal evolutionary strategies. We present four evolutionary strategies: Dominant, Altruistic, Balanced and Selfish and show how a linear combination can be employed in a fully self-consistent evolutionary game. Investigating fitness and evolutionary stable equilibriums, Genetic AI helps solving optimization problems with a set of predefined, discrete solutions that change dynamically. We show the universality of the approach on two decision problems.