Creative AI Through Evolutionary Computation: Principles and Examples

Abstract

The main power of artificial intelligence is not in modeling what we already know, but in creating solutions that are new. Such solutions exist in extremely large, high-dimensional, and complex search spaces. Population-based search techniques, i.e. variants of evolutionary computation, are well suited to finding them. These techniques make it possible to find creative solutions to practical problems in the real world, making creative AI through evolutionary computation the likely "next deep learning."

Keywords: Decision making; Evolutionary computation; Machine creativity; Neuroevolution; Real-world applications; Surrogate modeling.

Fig. 1

Challenge of Creative Problem Solving. Human design process as well as deep learning and reinforcement learning can be seen as hill-climbing processes. They work well as long as the search space is relatively small, low-dimensional, and well behaved. However, creative problems where solutions are not known may require search in a large, high-dimensional space with many local optima. Population-based search through evolutionary computation is well-suited for such problems: it discovers and utilizes partial solutions, searches along multiple objectives, and novelty. (Image credit: http://deap.readthedocs.io/en/latest/api/benchmarks.html)

Fig. 2

A comparison of human design and evolutionary design for a sign-up widget in web design. a The original design is clear and consistent, according to general design principles. b The evolutionary design is brash and bold, and unlikely to be designed by humans. However, it converts 45% better, demonstrating that evolution can discover creative solutions that humans miss

Fig. 3

Discovering a counterintuitive 24-h light period for computer-controlled agriculture. With the initial 18-h restriction removed, evolution discovered that when the lights are always on, basil will develop more flavor. The axes represent the three light variables, with light period on the horizontal axis. The color of the small dots indicates their value predicted by the model (red > yellow > green > blue). The large dots are suggestions, and the darker dots are the most recent ones. In this manner, if given a search space free of human biases, evolution can discover effective, surprising solutions

Fig. 4

An example creative solution for opening the economy after the COVID-19 peak had passed. The top plot shows the historical past and predicted future number of cases in the US on May 18th, 2020. The bottom plot illustrates the NPIs in effect or recommended during the same timeline, with color coding indicating their stringency. The system still recommends restrictions on schools, workplaces, and public events (top three rows), but suggests that opening and closing workplaces can be alternated, thus mitigating the effect on both economy and cases