Nature-Inspired Computation: two cases

Abstract

This work contains the result of two research lines that have ended up being submitted (one of them accepted and the other one is still under revision) to journals. On the one hand, Optimizing neural networks architectures with PBIL proposes a methodology to optimize the hyperparameters of the Inception-A block of Inception network. The main contribution is a special codification for the individuals being evolve that allows to skip the creation (hence, the evaluation) of non-valid individuals based on the requirements of this problem. Thus, this methodology is aware of the carbon footprint produced by many AI applications. In order to train and validate models, MNIST dataset is used. Results show this methodology can generate high-quality hyperparameters without explicitly search the complete space defined by the hyperparameters. On the other hand, Evolutionary Game Theory in a Cell: A Membrane Computing Approach proposes a general way to encode Evolutionary Game Theory into Membrane Computing and a novel computational framework which can be used to study, analyze and simulate the spreading of behaviours in structured populations organized in communicating compartments. In order to test the framework, two classic EGT games are used: Prisoner’s dilemma and Snowdrift game. The proposed approach allows to simulate populations organized in different compartments, allowing the study of the dynamics of populations that interact with each other. Results shows the spreading of behaviours in three cases: Prisoner’s dilemma, Snowdrift game and both games encoded in different membranes. As expected, results show that different behaviours (cooperators and defectors) can co-exists in the Snowdrift game, while in the Prisoner’s dilemma, the population is mostly composed by defectors.