Neuroevolution: Real-Time Creation of Sequential Digital Systems for Control, Design, and Decision Making
Computing & Wireless : Utility Software
Available for licensing
- Risto Miikkulainen, Ph.D. , Computer Science
- Kenneth Stanley, Ph.D. , Computer Science
Many applications lack a method of evolving networks for learning tasks. In difficult real-world learning tasks such as controlling robots, playing games, or pursuing or evading an enemy, there are no direct targets that can specify the correct action for every situation. Neural networks promise to solve this problem, but a long process of training must occur before it can be effective. A new system for the training of neural networks is needed and can be used to solve a number of problems.
Real-time NeuroEvolution of Augmenting Topologies (rtNEAT) is a genetic algorithm that trains and evolves neural networks of increasing complexity from a minimal starting point. This means networks that succeed continue while others are discarded, avoiding the problem of preparatory (non-real-time) training. Agents governed by rtNEAT neural networks can learn processes and even invent new solutions based on feedback without the guidance of a human programmer or controller, freeing the programmer from having to script extensive behaviors.
Can find solutions efficiently in real-timeCan solve new problems without trainingCan discover novel solutionsEvolves increasingly optimal and complex controllersCan be universally installed in systemsBroad range of beneficial applications
Continual, indefinite evolution Evolution occurs in real-time rather than at fixed intervals while the user has to waitBehavioral responses to environment and scenariosPackaged as a software development kit
Since NEAT and rtNEAT are general algorithms for evolving controllers, any application involving the automated control of some process, object, vehicle or sensory system could be viable. This technology currently is being directed towards the video game industry for the possibility of evolving characters in games and massive multiplayer online games. The uses for this algorithm, however, can be expanded to military simulations, educational games and applications, robotics, vehicle control systems, factories or as a research tool for modeling. The algorithms could also be implemented in pattern recognition and prediction applications.
Beta product/commercial prototype
- 1 U.S. patent issued: 7,559,843