Generative Design in Civil Engineering Using Cellular Automata Rafal Kicinger June 16, 2006
NKS 2006, June 16-18, 2006, Washington, DC 2 Outline Generative Design Cellular Automata as Design Generators –Steel Structures in Tall Buildings –Traffic Control Systems in Urban Areas Emergent Designer Design Experiments Experimental Results Conclusions
NKS 2006, June 16-18, 2006, Washington, DC 3 Generative Design: Representation Design representations –One of the key aspects of any computational design activity –Describe designs form, its components, etc. –Incorporate domain-specific knowledge –Determine the space in which solutions are sought Need to address important engineering objectives –Novelty –Optimization
NKS 2006, June 16-18, 2006, Washington, DC 4 Traditional Design Representations
NKS 2006, June 16-18, 2006, Washington, DC 5 Generative Design
NKS 2006, June 16-18, 2006, Washington, DC 6 Generative Design Cellular automata generating designs –Steel structural systems in tall buildings –Traffic control system in urban areas Evolutionary algorithms searching the spaces of generative representations (design embryos + design rules)
NKS 2006, June 16-18, 2006, Washington, DC 7 Cellular Automata as Design Generators Steel Structural Systems in Tall Buildings
NKS 2006, June 16-18, 2006, Washington, DC 8 Cellular Automata as Design Generators Traffic Control Systems in Urban Areas
NKS 2006, June 16-18, 2006, Washington, DC 9 Cellular Automata as Design Generators Traffic Control Systems in Urban Areas
NKS 2006, June 16-18, 2006, Washington, DC 10 Emergent Designer
NKS 2006, June 16-18, 2006, Washington, DC 11 Emergent Designer System architecture
NKS 2006, June 16-18, 2006, Washington, DC 12 Design Experiments Extensive Computational Experiments Conducted –Steel Structural Systems in Tall Buildings Exhaustive search of all elementary CAs started from arbitrary and randomly generated design embryos Generative representations based on 1D CAs evolved using evolutionary algorithms –Traffic Control Systems in Urban Areas Generative representations based on 2D CAs evolved using evolutionary algorithms
NKS 2006, June 16-18, 2006, Washington, DC 13 Design Experiments Steel structural systems: –number of bays - 5 –number of stories - 30 –bay width - 20 feet –story height - 14 feet Arbitrary design embryos used:
NKS 2006, June 16-18, 2006, Washington, DC 14 Design Experiments Traffic Control Systems –Number of network nodes- 65 –Number of network links -80 –Number of traffic signals - 25
NKS 2006, June 16-18, 2006, Washington, DC 15 Design Experiments CA representation parameters: –CA dimension: 1D and 2D –CA neighborhood radius: 1 and 2 –number of cell state values: 2 and 7 –CA neighborhood shape (2D CAs):Moore –CA iteration steps (2D CAs): 14 Evolutionary computation parameters: –evolutionary algorithm: ES –population sizes (parent, offspring): (1,5), (5,25),(5,125) –mutation rate: 0.025, 0.05, 0.1, 0.3 –crossover (type, rate): uniform, 0.2 –fitness: weight of the steel skeleton structure, or the total vehicle time
NKS 2006, June 16-18, 2006, Washington, DC 16 Experimental Results Exhaustive Search: Arbitrary Design Embryos Best designs:Total weight: Max. displacement:
NKS 2006, June 16-18, 2006, Washington, DC 17 Experimental Results Distributions plotted with respect to two objectives:
NKS 2006, June 16-18, 2006, Washington, DC 18 Experimental Results Exhaustive Search: Random Design Embryos Simple X bracings K bracings
NKS 2006, June 16-18, 2006, Washington, DC 19 Experimental Results Evolutionary search of generative representations: steel structures
NKS 2006, June 16-18, 2006, Washington, DC 20 Experimental Results Evolutionary search of generative representations: traffic control systems
NKS 2006, June 16-18, 2006, Washington, DC 21 Conclusions Generative representations based on cellular automata proved to perform well for civil engineering problems where some regularity/patterns are expected, or desired They produced quantitatively better solutions (6-20% average performance improvement) than traditional design representations
NKS 2006, June 16-18, 2006, Washington, DC 22 Conclusions CA representations produced qualitatively different patterns than patterns obtained using traditional representations They can be efficiently optimized by evolutionary algorithms, particularly in the case of 1D CA representations
NKS 2006, June 16-18, 2006, Washington, DC 23 Credits The work on generative design of steel structural systems in tall buildings was conducted together with Drs. Tomasz Arciszewski and Kenneth De Jong The work on generative design of traffic control systems in urban areas was conducted with Dr. Michael Bronzini
NKS 2006, June 16-18, 2006, Washington, DC 24 Backup Slides Evolutionary search of elementary CAs: K bracings
NKS 2006, June 16-18, 2006, Washington, DC 25 Backup Slides Evolutionary search of elementary CAs: Simple X bracings