Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Trackable Island-model Genetic Algorithms at Wafer Scale

Matthew Andres Moreno, Connor Yang, Emily Dolson, Luis Zaman

TL;DR

This work tackles the observability challenge in wafer-scale evolutionary computation by introducing a tracking-enabled asynchronous island-model GA designed for the Cerebras CS-2 Wafer-Scale Engine. It leverages hereditary stratigraphy markers to enable post hoc phylogenetic reconstruction and phylo-metrics, enabling insightful analysis of large-scale digital evolution at unprecedented throughput. The approach demonstrates on-hardware proof-of-concept with high generation rates and scalable population sizes, and shows that phylometric signals can differentiate adaptive from purifying dynamics, highlighting the method's potential for scalable, observable evolutionary research. The open-source kernel provides a drop-in framework for fixed-length genomes and fitness criteria, broadening applicability across the community.

Abstract

Emerging ML/AI hardware accelerators, like the 850,000 processor Cerebras Wafer-Scale Engine (WSE), hold great promise to scale up the capabilities of evolutionary computation. However, challenges remain in maintaining visibility into underlying evolutionary processes while efficiently utilizing these platforms' large processor counts. Here, we focus on the problem of extracting phylogenetic information from digital evolution on the WSE platform. We present a tracking-enabled asynchronous island-based genetic algorithm (GA) framework for WSE hardware. Emulated and on-hardware GA benchmarks with a simple tracking-enabled agent model clock upwards of 1 million generations a minute for population sizes reaching 16 million. This pace enables quadrillions of evaluations a day. We validate phylogenetic reconstructions from these trials and demonstrate their suitability for inference of underlying evolutionary conditions. In particular, we demonstrate extraction of clear phylometric signals that differentiate wafer-scale runs with adaptive dynamics enabled versus disabled. Together, these benchmark and validation trials reflect strong potential for highly scalable evolutionary computation that is both efficient and observable. Kernel code implementing the island-model GA supports drop-in customization to support any fixed-length genome content and fitness criteria, allowing it to be leveraged to advance research interests across the community.

Trackable Island-model Genetic Algorithms at Wafer Scale

TL;DR

This work tackles the observability challenge in wafer-scale evolutionary computation by introducing a tracking-enabled asynchronous island-model GA designed for the Cerebras CS-2 Wafer-Scale Engine. It leverages hereditary stratigraphy markers to enable post hoc phylogenetic reconstruction and phylo-metrics, enabling insightful analysis of large-scale digital evolution at unprecedented throughput. The approach demonstrates on-hardware proof-of-concept with high generation rates and scalable population sizes, and shows that phylometric signals can differentiate adaptive from purifying dynamics, highlighting the method's potential for scalable, observable evolutionary research. The open-source kernel provides a drop-in framework for fixed-length genomes and fitness criteria, broadening applicability across the community.

Abstract

Emerging ML/AI hardware accelerators, like the 850,000 processor Cerebras Wafer-Scale Engine (WSE), hold great promise to scale up the capabilities of evolutionary computation. However, challenges remain in maintaining visibility into underlying evolutionary processes while efficiently utilizing these platforms' large processor counts. Here, we focus on the problem of extracting phylogenetic information from digital evolution on the WSE platform. We present a tracking-enabled asynchronous island-based genetic algorithm (GA) framework for WSE hardware. Emulated and on-hardware GA benchmarks with a simple tracking-enabled agent model clock upwards of 1 million generations a minute for population sizes reaching 16 million. This pace enables quadrillions of evaluations a day. We validate phylogenetic reconstructions from these trials and demonstrate their suitability for inference of underlying evolutionary conditions. In particular, we demonstrate extraction of clear phylometric signals that differentiate wafer-scale runs with adaptive dynamics enabled versus disabled. Together, these benchmark and validation trials reflect strong potential for highly scalable evolutionary computation that is both efficient and observable. Kernel code implementing the island-model GA supports drop-in customization to support any fixed-length genome content and fitness criteria, allowing it to be leveraged to advance research interests across the community.
Paper Structure (3 sections, 2 figures)

This paper contains 3 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Results and Discussion

Figures (2)

  • Figure 1: Strategy for trackable distributed evolution simulation.Hstrat markers attached to agents support post hoc estimation of relatedness between lineages, enabling approximate phylogenetic reconstruction.
  • Figure 2: On-hardware Trial.Results from phylogenetic reconstruction of 1 million generation on-hardware simulations. Panel \ref{['fig:on-device-phylometrics']} compares phylometric readings from purifying-only and adaptation-enabled configurations, 4 replicates each. Panels \ref{['fig:on-device-adaptive', 'fig:on-device-purifying']} juxtapose example $750\times750$ PE simulation phylogenies under each configuration regime. Phylometrics were calculated from reconstructions with 10k sampled end-state agents. For legibility, phylogeny visualizations were further subsampled to 1k end-state agents. Top phylogenies are linear-scaled. Bottom phylogenies are log-scaled with ultrametric correction to better show topology.