Taking it to the next level: a new ‘multivalued’ neural net to identify mechanical designs
We designed an advanced ‘multivalued’ neural network to solve the classic kinematic chain isomorphism problem. It converges rapidly, is fully automatic, and requires no parameter tuning, outperforming previous approaches.
A few years ago, we successfully used a Hopfield neural network to tackle a classic problem in mechanical engineering: “are these two complex mechanisms really the same?” (the isomorphism problem). It was a good solution, but we knew we could do even better.
In our 2010 follow-up paper, published in the Journal of Computing and Information Science in Engineering, we presented the next evolution of that idea: a novel multivalued neural network that is faster, smarter, and easier to use.
🧐 The problem: seeking perfection
Our original Hopfield network was effective, but the quest for better “intelligent systems” in engineering is relentless. We wanted a method that was not just accurate, but also one that converged to a solution rapidly, could be fully automated, and didn’t require engineers to waste time “tuning” endless parameters to get it to work.
💡 Our solution: a multivalued neural network
Instead of a standard binary Hopfield network (where neurons are just ‘on’ or ‘off’), we designed a multivalued neural network. This new architecture allows us to formulate the graph isomorphism problem in a much simpler and more elegant way.
But we didn’t stop there. We improved the model’s performance by adding an “extra constraint” that forces the network to also consider the degree of the vertices (how many links each joint has). This simple addition dramatically speeds up the search for a solution.
🚀 The key advantages: fast, automatic, no tuning
This new model proved to be a significant leap forward. Its three main advantages are:
- It converges rapidly: The new design and the added constraint help the network “settle” on the correct answer extremely quickly.
- It needs no parameter tuning: This is a huge benefit. The algorithm works out-of-the-box without needing an expert to tweak its internal settings.
- It’s fully automatic: The entire process is designed to be an automated tool for intelligent design and manufacturing systems.
We tested our new multivalued neural network against other recently presented approaches, and the simulation results were clear: our model performed better and faster.
🔬 Why does this matter?
This work provides engineers with a state-of-the-art computational tool. By creating a faster, more reliable, and “zero-config” algorithm, we make it practical to build smarter design software that can instantly check new mechanisms against vast libraries of existing ones, accelerating innovation in robotics and mechanical design.
📖 The full paper
For the complete technical breakdown of the multivalued network model, the mathematical formulation of the new constraints, and the simulation results, you can read the full journal article.
A new multivalued neural network for isomorphism identification of kinematic chains. Authors: Gloria Galán Marín, Domingo López-Rodríguez, Enrique Mérida Casermeiro. Journal: Journal of Computing and Information Science in Engineering (vol 10, issue 1)