Evolution of Vision-Based Agent Behaviour in Hilly LandscapesAris Alissandrakis Kerstin DautenhahnDepartment of CyberneticsUniversity of ReadingPO Box 225 ReadingRG6 6AY, UKshu95aa@reading.ac.uk K.Dautenhahn@reading.ac.ukAbstractThis paper presents research in the field of sensor evolution, using computer simulations of agents moving in a hilly landscape, with evolving vision sensors. The model of the environment and the rules for the agents are presented, along with the results of a few experiments implemented using the SWARM Simulation System. The initial results are discussed and further developments of the project are considered.1 INTRODUCTIONBiological evolution shows impressive examples of how sensors are adapted 1) to the specific constraints of the habitat which an animal is living in, and 2) to the survival strategy of the animal population in this environment. Studying sensor evolution with artificial agents is therefore a challenging research area, compare (Todd et al 1993, Menczer et al 1994, Mark et al 1998). For the study of sensor evolution a useful and feasible simulation model of qualitative sensor evolution must be constructed. Having obtained the model, one can then identify the mechanisms that lead to new sensor generations. The current paper describes such a model and presents some initial experimental results. The simulations study 2 1/2 dimensional hilly landscape environments. Previous experiments with mobile robots in hilly landscapes are described e.g. in (Dautenhahn 1995, Billard and Dautenhahn 1997).2 THE MODELThis section describes the environment and the agents for the world model used in the experiments. The agents live in a hilly environment, and move around using vision sensors that are evolved through successive generations, by a simple genetic algorithm.2.1 THE ENVIRONMENTThe environment consists of a 2-dimensional grid, each cell containing positive integer values. The values represent the altitude of the landscape, and therefore the spatial and numeric distribution is important. The numeric difference between nearby cells is kept small in order to avoid large discontinuities and to have a smooth resulting surface. The distribution must be seamless - i.e. the opposite edge values must correspond without large gaps - so that the agents can move on a torus surface (see Fig. 1).Figure 1: An example landscape containing four valley areas, two of them connected by a canyon. The landscape remains unchanged during each experiment. The environment does not contain any food or energy resources for the agent. Each grid position is only able to contain a single agent and each agent (if not completely surrounded by other agents) has to move to a different grid position in each time step.
本文来源：https://www.2haoxitong.net/k/doc/7b93b02db4daa58da0114ab6.html
