Six Generative Design Pitfalls to Avoid – Part Two

Who Needs Designers?, Paradox of Choice, and Designs Not Metrics

Peter Bentley


In part one of my article, I shared the first of three generative design pitfalls to avoid. Today, I’m sharing three more, using my personal experience and three decades of experience in generative design.

Four: Who Needs Designers?

Many researchers who tackle generative design come from a computer science or mathematics background. I’m one of them. It makes us very good at creating clever algorithms, but very bad at understanding what the goal of research should be. When I started in this area I fell into the same trap as everyone.

“Component-based representations enable the system to assemble entirely novel forms, like a child building a design from LEGO bricks,” I explained. “This is quite different from optimizing parameterized designs, where all you can do is tweak an existing solution. By starting with random collections of components and letting the system choose the number and configuration of the components, you can have a complete design system. It creates the design from nothing and optimises it until it meets the objectives perfectly. Any questions?”

The auditorium was suddenly filled with the sound of fists banging on desks and feet stomping on the wooden floor. I looked around in surprise. Were they happy or upset?

The first time I heard German-style applause was after giving a talk at DaimlerChrysler more than 20 years ago. All that foot-stomping and fist-banging scared the life out of me. Apparently, it’s a good thing. But despite the positive reception, the automobile designers were terrified of my vision for generative design. They didn’t want their jobs taken away from them by a computer.

This was the start of my education about what the world really wanted. As a computer scientist I’d always pushed to make my systems automate as much of the design process as I could. Who needs designers at all? A system that could design everything by itself was surely the ultimate achievement. It’s a view I still hear today from computer scientists. But in the real world, I came to realise we always needed the expertise of designers with their vast experience and domain knowledge.

It is hopelessly naïve to believe we’re going to have humanoid robots that will replace our physical labour in the next few decades. It is equally naïve to believe that a computer can replace an experienced designer any time soon. But most importantly – even if we could do this, nobody wants this. Part of being human is to design and make. Why should we make ourselves less human? Generative design should not mean automated design. Generative design should mean helping humans be better designers. We all need to remember this lesson in our research.

Five: Paradox of Choice

Replacing designers won’t work. But the next pitfall is the opposite: we overcompensate and try to help designers too much, by giving them too much choice.

I would like “Fugue” to be about the immune system,” said new media artist Gordana Novakovic. “Immune cells, blood cells should be visible from within the body. The audience should affect the behaviour of the artwork, like chemical emotions flowing through.” She looked at me. “Can you do it?”

“Sure,” I said. “We can make perfectly realistic looking cells so that participants will see the body as it really is.”

“I don’t want that,” said Gordana. “I want my artistic intent expressed in the design of the cells.”

“We can let you generate your own forms in an evolutionary art system and then you can have any shape that you like,” I countered.

“So I would have to look at hundreds or thousands of options?” Gordana paused. “I’ll just make them out of clay.”

Giving the designers choice is great. Giving them so much choice that they have to hunt for acceptable designs like a needle in a haystack is not so good. Fugue was a fairly typical collaboration where the very different mind of the artist was enabled by our scientific and engineering skills. In a sense we became the brushes and paints for Gordana as she imagined her artwork with us. And while we were happy to provide her as much choice as she wanted, often she would revert to the tools she knew, such as model-making with clay.

Tony Ruto, Director of Research Engineering at Autodesk, helped by 3D-scanning the clay models of cells produced by the artist and we used one of my immune algorithms combined with a swarming algorithm that resulted in the aggregation of simulated platelets in a clot. Together Tony and I created a 3D “Fantastic Voyage”-style artwork where you were immersed within the blood vessels of a virtual body (real-time graphics projected onto curved screens), watching and affecting the immune reaction take place, with audio provided by an audio artist. Gordana toured the world with the artwork.

Designers need control. Excessive choice does not give control. Psychologists call it the Paradox of Choice: too much choice can result in poor decision-making. Over the years I’ve seen entire fields of research become distracted by a perceived need for choice: multi-objective optimisation focusses entirely on the idea that finding an even coverage of solutions on a Pareto Front is the goal. Once you have these, you “just” show them all to a designer, who can then assess which compromises they prefer. Such researchers focus all their efforts on achieving the first goal but never see how their methods are received by users. When the number of objectives becomes more realistic (usually considerably more than three or four), the number of options can become overwhelming.

It’s not the number of options that matters, it’s their quality. When we give designers choice, the choice should be between a small number of genuinely viable options. Control is even better: give a designer a small number of parameters they can modify such that the designs can be varied in meaningful ways, with all variations valid and viable options.

Six: Designs not Metrics

The final pitfall that I’ll mention here is one that I see all too often in research papers. The paper usually looks like this (and I’ll paraphrase a little):

We’ve created a new generative method by mixing a few existing ideas together plus adding something new. We tested it on an established/difficult problem/dataset. Our results are x% according to the metric, which is better than current state of the art.

I look through the paper to see examples of the actual solutions generated by the system. Nothing. The paper reports only the metric. When I meet the authors, I ask: “so what did the results actually look like?”. “We only looked at the metric,” they reply.

Over the years I’ve been privileged to work with a lot of different talented PhD students, exploring different aspects of generative design. Each project was several years of investigation. Guess what? We always looked at what the system generated!

Prototype 3D printed robot snake by Siavash Mahdavi

Sanjeev Kumar studied developmental processes and morphogenesis, with wonderful support from biologist Lewis Wolpert. His system produced some of the earliest evo-devo forms. They often looked like weird blobs.David Basanta studied the evolution of microstructures with the support of material scientist Mark Miodownik. His system produced crystalline structures. (David now runs his own lab studying the growth of cancer through computational modelling.)

Tim Gordon generated novel electronic circuits in the new field of evolvable hardware. His system produced some of the largest computer-generated adder circuits to date, making use of repeated modules in its design.

Katie Bentley (no relation) studied the remarkable morphogenesis in natural diatoms. Her system produced pretty fractal-like forms. (She now also runs multiple labs on modelling of vascular systems in cancer.)

Boonserm Kaewkamnerdpong studied swarm-like self-assembly through aggregation of particles. Her system produced 2D patterns and even demonstrated virtual blood vessel repair.

Navneet Bhalla studied 3D-printed self-assembling systems. His system produced 3D-printed forms that when agitated, self-assembled into desired structures.

Hooman Shayani studied evo-devo neural networks in hardware. His system generated a spiking neural Cortex microcircuit.

Siavash Mahdavi studied the evolution of robot snakes and their internal microstructure. He didn’t just generate robot snakes, he ended up making a company, hiring Hooman and Tony, and after a lot of hard work by them all, having it acquired by Autodesk. But that’s another story…

In every one of these generative design projects (and many others) the repeated lesson I have learned is this: look at the designs and not just the metrics. Sometimes (quite often) you might find that the best design produced according to the metric is nothing more than a cheat, exploiting some loophole in the representation, algorithm or metric. Sometimes you find that terrible designs according to the metric show a bizarre stroke of genius.

Generative design is an amazingly fun area to work in because of what it can generate. Both for the researchers and the designers, what matters is the design. Yes, the metrics are very important. But equally important is the surprise at seeing something that no-one has ever seen before. That brand new creation suggested by the system. Maybe it is a bit silly. Maybe it needs some revision and tweaking. But generative design can be about novelty and creativity as much as perfection. Sometimes a design that breaks all the rules is the one that the designer is looking for. They just didn’t know it until they saw it.

Today is a new time of excitement that reminds me of the era I began in this field 30 years ago. We have extraordinary new machine learning algorithms emerging that use the vast data and computation available today to generate new designs from little more than a textual description. Combine these with some well-polished techniques that we have taken to maturity and we will have a new suite of generative design techniques.

The problems of generative design are not solved today. I do not think they will be solved tomorrow. Working together we can help our future designers reach their full potential. Not by replacing them – I learned that lesson a long time ago. By helping them understand new possibilities in design, materials, analysis and by giving them intuitive new tools to help them make our world better.

In case you missed it, here are three additional generative design pitfalls to avoid in Part 1 of this post.

Peter Bentley is a Visiting Professor at Autodesk and Honorary Professor at University College London.

Get in touch

Have we piqued your interest? Get in touch if you’d like to learn more about Autodesk Research, our projects, people, and potential collaboration opportunities

Contact us