Posted on 28 April 2020
Their article, entitled 'Novelty Search for Shape Descriptors', looks at a new way of approaching artificial evolution to develop physical shapes.
Artificial evolutionary design is a method which takes inspiration from real-world biological evolution. By using evolutionary algorithms, innovative designs can be developed without being heavily influenced by previous products. Innovation in product design is crucial, not only to ensure market competitiveness, but to meet increasingly stringent engineering and manufacturing requirements. However, the use of these evolutionary algorithms for generative design still remains challenging due to an innate restrictiveness in their operation.
Dr Hickinbotham and Professor Tyrrell's work, in collaboration with researchers at the Queen’s University Belfast, combines current evolutionary design methods with a Novelty Search functionality, looking at determining how 'evolvable' initial designs can be. In their work, they have replaced the part of the evolutionary algorithm which determines how well a design fits specification, with a function that instead looks at how novel a design is compared to previous iterations. By doing this, they found that designs could be pushed much further than simple random mutations would, giving them a benchmark for how 'evolvable' initial structures can be.
Future work for this project hopes to focus on methods for adding environmental feedback to the algorithm, such as mechanical stress, and including specific measures to drive the later stages of evolution towards a range of manufacturable solutions.