Distributed Hardware Evolution Project

Distributed Hardware Evolution Project
Project
Status	Discontinued
Category	Evolutionary Computing
Compute	CPU
Requires	None
Development
Developer	Michael Garvie
Author	Michael Garvie
Sponsor	University of Sussex
Maintainer	None (project discontinued)
Initial release	May 11, 2018  (8 years ago)
Discontinued	Autumn 2019
Software
Written in	Java (client)
Operating system	Windows, Linux, macOS
Metadata
Website	http://dhep.ga/boinc/

DHEP (the Distributed Hardware Evolution Project) was a BOINC-adjacent volunteer computing project based in the Department of Informatics at the University of Sussex, England. The project allowed volunteers to host an "island" running a population-based metaheuristic stochastic optimisation algorithm – specifically a genetic algorithm – in a coevolutionary setting, with the aim of automatically synthesising future super-reliable electronics of the kind used in autonomous vehicles, power stations, medical equipment, and aerospace systems. As increasing numbers of human lives depend on correctly functioning hardware, the project framed its work as directly relevant to safety-critical engineering, while also using the resulting data on population dynamics to study migration rates, genetic diversity, and the mechanisms of genetic recombination more generally.

The project was administered by Michael Garvie, a University of Sussex researcher who had studied evolutionary approaches to self-checking hardware for over a decade before DHEP's launch. DHEP entered public beta on 11 May 2018 and was suspended in autumn 2019 pending funding that never materialised.

Falmer House at the University of Sussex, where DHEP was based within the Department of Informatics.

History

DHEP grew directly out of Garvie's doctoral research at Sussex. His 2005 PhD thesis, Reliable Electronics through Artificial Evolution, and an earlier paper co-authored with Adrian Thompson, "Evolution of Self-diagnosing Hardware" (2003), had already demonstrated that evolutionary algorithms could design small self-testing circuits. DHEP was conceived as a way to scale that earlier, small-scale work up to a distributed grid of volunteer processors, allowing much larger and more numerous "islands" of evolving circuit populations to run in parallel than would be feasible on Sussex's own hardware alone.

A forum poster who recalled an earlier, pre-BOINC incarnation of the project noted that it had folded once its original student contributor, referred to informally as "Miguel," finished his time at university, with the project restarting later once BOINC integration was implemented. The BOINC-integrated version was announced and began exporting statistics to third-party BOINC statistics sites during 2018, at which point the project entered what its team described as a "production" state.

DHEP was shut down suddenly in autumn 2019. In a message to volunteers, the project's administrator, signing as "Michael," explained that DHEP had to be suspended until further funding could be secured, thanking the BOINC community and volunteers for their contributions.

The evolutionary methodology

DHEP's underlying science goal was concurrent error detection (CED): adding hardware logic to a digital circuit so that it can signal an error as soon as one occurs, without waiting for external testing. A circuit with this property for a given fault set is described as totally self-checking (TSC) if it is both self-testing (every fault in the set is eventually signalled during normal operation) and fault-secure (no fault ever produces an incorrect output without an accompanying error signal).

A field-programmable gate array die. DHEP's genetic algorithm evolved digital logic circuit designs of this general kind, targeting a generic two-input-gate technology library.

Rather than using conventional, hand-designed CED techniques such as duplication-and-comparison or parity checking, DHEP searched for TSC circuit designs directly, using a genetic algorithm operating on populations of candidate circuits encoded as binary genotypes. Each individual in a population represented a complete circuit design, instantiated and evaluated in a digital logic simulator against several fitness criteria simultaneously: whether it computed the correct output function, whether it was self-testing, whether it was fault-secure, and how small (parsimonious) the resulting circuit was.

For the output-function component of fitness, each candidate circuit's response $Q_j$ at output $j$ was compared with the desired response ${Q_j}^{\prime }$ across all output lines $q$, giving an overall function fitness

$f_f=\sum _{j=0}^{q-1}\frac{|corr(Q_j,{Q_j}^{\prime })|}{q}$

with $f_f=1$ indicating a circuit that reproduces the target function exactly. Self-testing and fault-secure fitness were evaluated by simulating the circuit under every fault in a defined fault set and measuring how many faults or fault/input-word combinations failed to be correctly signalled, penalising unsignalled faults according to

$f_{ST}=\frac{1}{1+u_f{k}_{ST}}$ and $f_{FS}=\frac{1}{1+u_i{k}_{FS}}$

where $u_f$ and $u_i$ count unsignalled faults and fault/input-word instances respectively. A circuit scoring the maximum on all three metrics was mathematically guaranteed to be TSC with respect to the fault set used.

Island-based coevolution

To exploit the volunteer computing grid, DHEP distributed this genetic algorithm using an island-based coevolutionary model. Each connected volunteer computer was assigned a location on a two-dimensional grid and ran its own independent, small, genetically converged population – an "island" – of 32 candidate circuit genotypes evolving over many generations, with mutation as the primary driver of variation. Individuals were periodically selected to migrate between islands, with the probability of a given destination island being chosen varying inversely with its grid distance from the source, allowing fit genetic material to diffuse across the whole population of islands while largely preserving genetic diversity within each one. Because islands could join or leave the grid as volunteer machines came online or dropped out, the algorithm was inherently tolerant of a dynamically changing number of participating hosts – a natural fit for a volunteer computing model. At its largest, the system described in the project's associated research paper made use of a grid of up to 150 workstations.

Client and credit system

Because DHEP predated its BOINC integration, its computational core remained a bespoke, Java-based application communicating with DHEP's own servers rather than a conventional BOINC science application; the Windows/Linux/macOS client was distributed under the name "Distributed Hardware Evolution Island." BOINC-issued workunits functioned as long-running placeholder tasks of indefinite duration rather than discrete jobs to be completed: actual evolutionary progress was tracked server-side, and credit was granted through an hourly trickle-message system similar in spirit to that used by climateprediction.net. A given task would effectively run "forever" from the client's point of view, terminating only when the project's science goal was updated on the server, which volunteers reported happened every few weeks. DHEP tasks did not support checkpointing and offered no GPU, Android, Raspberry Pi, or NCI application variants.

Autonomous vehicles were cited by DHEP as an example application area for the super-reliable, self-checking hardware its algorithm aimed to design.

Applications

DHEP framed concurrent error detection as increasingly important wherever hardware failure carries a direct risk to human safety, explicitly citing autonomous vehicles, power stations, medical equipment, and aerospace systems as target domains for the kind of low-overhead, self-checking circuitry its evolutionary process was designed to discover. The project's own results paper reported that its evolved circuits achieved TSC concurrent error detection using an average of only 23% of the overhead required by conventional duplication-based checking across a standard set of benchmark circuits, compared with roughly 69% for the best previously published techniques.

Scientific publications

Publications using BOINC-computed data

The following publication explicitly credits computation donated by the distributed computing community, corresponding to processing performed on the DHEP volunteer grid, in addition to time on the Grid'5000 research testbed:

• (2019). "Automatic Synthesis of Totally Self-Checking Circuits". arXiv: 1901.07023.

Other project-related publications

These earlier works by DHEP's founder predate the project's 2018 launch and describe the precursor research from which DHEP's methodology was developed; they did not use BOINC- or DHEP-computed data:

• (2005).Reliable Electronics through Artificial Evolution. University of Sussex.
• Garvie, M..(2003})."Evolution of Self-diagnosing Hardware".Edited by A. Tyrrell, P. Haddow, J. Torresen.pp. 238–248.

• Garvie, M..(2003})."Evolution of Combinational and Sequential On-line Self-diagnosing Hardware".Edited by J. Lohn, R. Zebulum, J. Steincamp, D. Keymeulen, A. Stoica, M. Ferguson.pp. 167–173.

See also

• BOINC

External links

• http://dhep.ga/boinc/ (offline since project discontinuation)
• https://arxiv.org/abs/1901.07023
• [DHEP entry on the BC-Team wiki DHEP entry on the BC-Team wiki]