est. 2001
Applying genetic
algorithms to
neural networks.
Xtructure is an artificial intelligence research company. We develop algorithms, tools, and theory at the intersection of evolutionary computation and deep learning.
Research Focus
Evolutionary Computation
Genetic algorithms as a mechanism for optimizing neural network topology, weights, and learning rules — without gradient descent.
Discrete Event Simulation
High-fidelity simulation environments for testing evolutionary strategies across complex, event-driven systems.
Graph Data Structures
Novel hypergraph representations for encoding complex relational structure in AI knowledge systems.
Open Source · 2012
XNet
XNet is a Java platform for discrete event simulation and computational evolution — released open source in 2012. It has been used in production research for over a decade, and a Python port is currently underway.
The framework provides a composable toolkit for constructing evolutionary experiments: population management, genotype/phenotype encoding, fitness evaluation, and simulation scheduling — all in one cohesive library.
Patent · 2025
XBundle
XBundle is a hierarchical hypergraph database, patented in 2025. It is designed to represent and query complex, multi-relational data structures that exceed the expressiveness of conventional graph databases.
Where standard graph models connect nodes with binary edges, XBundle's hypergraph formalism allows edges to span arbitrary sets of nodes — enabling richer representations of structured knowledge, including those required by advanced AI systems.
Founder
Michael Roberts
Founder & Lead Researcher
Michael Roberts has two decades of experience in the field of artificial intelligence and machine learning. Along the way he has had the privilege of working with great teams at organizations like NASA, The Walt Disney Company and TomTom. His career in technology has progressed from the design of advanced systems for manned spaceflight in the days of the Space Shuttle, through building an experimental search engine in the early days of the Internet, to exploring the application of genetic algorithms to dynamic neural networks and hypergraphs.