Search

IEEE Computer Society Press.

Cambridge, MA: MIT press

The problem of the validity of simulation is particularly relevant for methodologies that use machine learning techniques to develop control systems for autonomous robots, as, for instance, the artificial life approach known as evolutionary robotics. In fact, although it has been demonstrated that training or evolving robots in real environments is possible, the number of trials needed to test the system discourages the use of physical robots during the training period. By evolving neural controllers for a Khepera robot in computer simulations and then transferring the agents obtained to the real environment we show that (a) an accurate model of a particular robot-environment dynamics can be built by sampling the real world through the sensors and the actuators of the robot; (b) the performance gap between the obtained behaviors in simulated and real environments may be significantly reduced by introducing a “conservative” form of noise; (c) if a decrease in performance is observed when the system is transferred to a real environment, successful and robust results can be obtained by continuing the evolutionary process in the real environment for a few generations.

In this article we describe a new approach in evolutionary robotics according to which human breeders are involved in the evolutionary process. While traditionally robots are selected to reproduce automatically according to a fi tness formula, which is a quantitative and strictly defi ned measure, human breeders can operate selection based on qualitative criteria, and rewarding behaviors that can slip between the meshes woven by the fi tness formula. In authors’ opinion this may bring advantages to the evolutionary robotics methodology, allowing the production of robots that display more, and more multiform, behaviors. In order to illustrate this approach, the software Breedbot was developed in which human breeders can intervene in evolving robots, complementing the automatic evaluation. After describing the software, some results on sample evolutionary processes are reported showing that the joint use of human and artifi cial selection on an exploration task generates robots with a higher performance and in a shorter time compared with the exclusive action of each breeding method. Future work will explore this hypothesis further.

The authors explore the concept of development without programming by children. Especially, they look at the case of developing robot control systems. The evolutionary robotics approach has shown that in some cases, given a mathematically described fitness function, it is possible to achieve an automatic development of robot controllers. However, it is questionable how one is to construct the mathematical fitness function. So they applied an interactive genetic algorithm to the problem of developing robot controllers and achieved and evolutionary robotics approach that allows children without any programming knowledge to develop controller for LEGO robots. They used neural networks as robot controllers, and found that combining the interactive genetic algorithm with a kind of reinforcement learning-development at the evolutionary time scale combined with life-time development-reduces the development time drastically. Hence, they overcome one of the major drawbacks of the interactive genetic algorithm, namely the development time.

Currently, evolutionary computation (EC) typically takes place in batch mode: algorithms are run autonomously, with the user providing little or no intervention or guidance. Although it is rarely possible to specify in advance, on the basis of EC theory, the optimal evolutionary algorithm for a particular problem, it seems likely that experienced EC practitioners possess considerable tacit knowledge of how evolutionary algorithms work. In situations such as this, computational steering (ongoing, informed user intervention in the execution of an otherwise autonomous computational process) has been profitably exploited to improve performance and generate insights into computational processes. In this short paper, prospects for the computational steering of evolutionary computation are assessed, and a prototype example of computational steering applied to a coevolutionary algorithm is presented.