The Dancing Salesman Problem

What does it mean for software to be truly, independently creative? With the Painting Fool project – described at www.thepaintingfool.com – our aim is for the software to be taken seriously as a creative artist in its own right, one day. For 25 years, we’ve been handing over creative responsibilities to The Painting Fool, within the AI research field of Computational Creativity. Through much scientific and artistic exploration, members of this research community have built up a philosophy around the notion of AI systems being genuinely creative, surpassing the basic requirement of producing artefacts of value. If it is to have any chance of earning the description ‘creative’, software needs to exhibit behaviours which reflect skill, appreciation, imagination, learning, accountability and intentionality in creative endeavours, innovating in technique, aesthetic judgment and imaginative reasoning. The Painting Fool has been a test-bed for implementing such behaviours and exploring the cultural impact of creative artificial intelligence, long before generative AI systems became commonplace. Overlapping historically with Harold Cohen’s AARON and Mario Klingemann’s Botto projects, The Painting Fool produces artworks in a relatable way, employing ideation processes to construct scenes which are then rendered using simulated art materials. Feedback from artistic projects with the software has led to avant-garde philosophical ideas, such as the “Machine Condition” framework: inspired by literature on the human condition, we argue that software should communicate through its art what it is like to be it, and what it is like to be an AI system in general. Interacting with artistic communities has always been part of the process of building The Painting Fool, and criticisms of its techniques and outputs have regularly led to changes in its code and advances in its sophistication. In the “You Can’t Know my Mind” exhibition in Paris, 2013, we addressed a perceived lack of intentionality in creative AI systems, by enabling The Painting Fool to paint portraits driven by a simulated mood gained through reading newspaper articles. In a terrible mood, the software told sitters for a portrait to go away, with an explanation of its reasons for doing so. In better moods, it used the sitter as a source material for portraiture, and its mood to determine the way in which it painted the portrait. When the portrait was finished, The Painting Fool used early neural machine vision techniques to determine whether the artwork reflected its mood or not, and learned from the results to be better. Afterwards, we would ask the sitter whether they felt the software intended to make the artwork in the way that it did. The Dancing Salesman Problem artworks were produced for an exhibition entitled “No Photos Harmed” in Paris, 2011, and for the 2014 “Creative Machines” exhibition at Goldsmiths College in London. The name reflects the classic computer science problem where a travelling salesman must drive from town to town without returning to one previously visited. Mapping towns onto colour regions and driving onto brush strokes, The Painting Fool produced these dynamic pieces with swoops representing large distances driven to find the next unvisited town. The figures were generated using AI techniques including context-free design grammars and constraint solving, showing that fully-automatically produced pieces can be representational rather than abstract, without requiring photographic input, hence no photos being harmed. By this stage in its development, The Painting Fool could simulate the use of acrylic paints, pastels, pencils, aerosols, felt-tip pens, chalks, charcoal and watercolours running into each other. Given the traveling salesman nature of the layout, each piece is produced in one long stroke, changing colour and tapering out where needed. At a higher level, the project explored the notion that AI systems like The Painting Fool could be imaginative by inventing scenes that never existed in reality. While GAN or text-2-image diffusion models can routinely do this now, in the 2010s there were only a handful of systems able to imagine worlds and render them as beautiful artworks.