An Evolvable Hardware Chip and Its Application as a Multi-Function Prosthetic Hand Controller

AAAI 1999

TL;DR

This paper introduces an "Evolvable Hardware" (EHW) chip that can autonomously reconfigure its own circuit structure using genetic algorithms. This innovation significantly speeds up the adaptation process for multi-function myoelectric prosthetic hands, reducing training time from weeks to minutes and paving the way for more practical and user-friendly prosthetics.

What Problem Does This Paper Address?

Controlling advanced prosthetic hands, particularly those with multiple functions, typically requires a lengthy and intensive training period for the user. This is because the myoelectric signals (EMG) generated by muscle movements vary significantly between individuals and can even change over time for the same person. Conventional approaches often rely on pre-programmed hardware or software simulations that are either inflexible or too large and power-hungry to be integrated into a prosthetic device. The challenge is to create a control system that can adapt quickly and efficiently to the user's unique EMG signals and the dynamic nature of these signals, all within the size and power constraints of a prosthetic limb.

Methods

The core innovation is a custom VLSI chip implementing "Evolvable Hardware" (EHW). EHW combines reconfigurable hardware (specifically, Programmable Logic Arrays or PLAs) with genetic algorithms (GAs). The chip's circuit configuration is represented as a "chromosome" (a bit string), and GAs are used to evolve this chromosome to achieve a desired function.

Key advancements in this paper include:

  • VLSI Implementation: Unlike previous EHW work done in software simulations, this paper presents a compact, on-chip implementation, making it suitable for embedded applications like prosthetics.
  • Gene Replacement Genetic Algorithm (GRGA): To accelerate adaptation, GRGA was developed. It allows for the direct replacement of parts of a chromosome with "chromosome candidate segments" derived from training input-output patterns. This is a significant speedup over traditional GA operations.
  • On-line Training Pattern Memory Edit Mode: This feature allows the training data stored on the chip to be edited dynamically during the genetic algorithm's operation. This is crucial for adapting to changing EMG signal characteristics in real-time.

The EHW chip's architecture includes a GA Unit, various memory blocks (for chromosomes, training patterns, and fitness values), two PLA units for parallel circuit evaluation, and a small CPU for interface and some fitness calculations.

Datasets

The paper describes experiments using training sample patterns for EMG signal classification. These patterns consist of input EMG signal features and corresponding desired prosthetic hand actions (e.g., open, grasp, flex, extend).

  • Initial Training: 60 patterns (10 patterns x 6 actions) were used.
  • Additional Training: 30 patterns were added for actions with lower initial performance.

The paper does not introduce a new, publicly available dataset but rather uses a custom-generated set of training patterns derived from simulated EMG signals and desired prosthetic hand actions.

Key Insights and Ideas

  • Hardware that Learns: The most significant insight is the feasibility of creating hardware that can autonomously reconfigure itself to perform a specific task. This moves beyond fixed-function hardware towards adaptive systems.
  • Speeding Up Adaptation: GRGA is a critical development, demonstrating that by intelligently incorporating information from training data directly into the genetic search, adaptation speed can be dramatically improved (over six times faster than previous methods in their simulations).
  • On-line Adaptability is Crucial: The "on-line edit mode" highlights the necessity of continuous adaptation for real-world applications like prosthetics, where signal characteristics are not static.
  • Compactness for Practicality: The VLSI implementation proves that complex adaptive systems can be miniaturized to fit within the constraints of wearable devices.

Key Results

The paper reports on simulations for a prosthetic hand controller. The EHW chip, using GRGA, achieved an average output pattern rate of 81.0% for EMG signal classification. This performance is comparable to neural network controllers that use on-line training approaches (reported at 81.5%).

Crucially, the GRGA approach required significantly fewer evaluations (approximately 9710 on average) to synthesize a basic combination circuit compared to standard GA operations (approximately 61874 on average). This translates to a much faster adaptation time, potentially reducing the user training period from weeks to about ten minutes.

Limitations and Future Work

  • Simulation-Based: The experiments were conducted using a simulator on a PC, not on the actual EHW chip in a real prosthetic hand. While promising, real-world deployment may reveal further challenges.
  • Limited Number of Actions: The experiments focused on classifying EMG signals for six prosthetic hand actions. More complex prosthetics with a wider range of movements might require more sophisticated EHW designs.
  • Feature Vector Simplification: The paper uses integrated EMG (IEMG) signals as feature vectors, which are simpler and less computationally intensive than frequency spectrum analysis. While this addresses the computational overhead problem, it might limit the precision or nuance of control compared to more complex feature extraction methods.

Future work could involve testing the EHW chip in actual prosthetic devices, exploring more complex EMG signal processing techniques, and expanding the number of controllable functions. The authors also mention plans for commercialization and adoption by Hokkaido University's medical department, indicating a strong belief in the technology's practical potential.

Summary updated

BibTeX
@inproceedings{1999evolvable,
  title = {{An Evolvable Hardware Chip and Its Application as a Multi-Function Prosthetic Hand Controller}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure