Intelligent Cooperative Multi-Agent-Systems (MAS)

Research project P 13937 Intelligent Cooperative Multi-Agent-Systems Peter KOPACEK 11.10.1999 Approximately 15 years ago the keyword "Multi-Agent- Systems (MAS)" was introduced. MAS have emerged as a sub-field of Al and mostly applied in software engineering. in this particular case intelligent agents are dealing with software. packages, which are able to solve the sub-tasks autonomously. For this purpose agents should be intelligent, autonomous, cooperative and communicable. Parallel to new developments in factory automation (Intelligent Manufacturing Systems (IMS), agile manufacturing) there are approaches to implement the gained knowledge from MAS in these fields. In this case the agents are mobile robots (hardware agents), which are able to solve a global task together, In many industrial applications not only one mobile robot but several ones are working together. But comparing to MAS these robots are not able to define the tasks by themselves. Robots (Agents) are working in a dynamically changing environment consisting of fixed and movable obstacles. The latter can be influential (other agents) and not influential (other movable obstacles). In order to influence the behavior of other agents the agent should have ongoing interaction with this environment and with other agents. The type of interaction, such as exchanging data and information, depends on the complexity of the system. Based on the results of the previous project, the goal of this project is the development of control algorithms for intelligent and autonomous group behavior of robots (agents) mainly in conflict situations. For this purpose, two problems to be solved are the communication as the basis for learning behavior between agents and the development of interaction patterns combinations of cooperative and coordinated interactions - using AI-Methods (artificial neural networks and neurofuzzy). Neural networks and neurofuzzy-based MAS-algorithms developed in this project will be implemented and tested on a group of homogeneous and heterogeneous mobile robots (hardware agents) for possible industrial application in small and medium sized enterprises (SME`s), such as transportation tasks in a factory. This research work will be accomplished in the framework of a cooperation agreement between FWF (Austrian Science Fund) and KOSEF (Korea Science and Engineering Foundation) signed on November 28, 1998. The Institute for Handling Devices and Robotics (IHRT), the Vienna University of Technology and the Korea Advanced Institute for Science and Technology (KAIST) are project partners for a cooperative research.

This research project dealt with the navigation of mobile intelligent robots. In contrast to stationary unintelligent robots the navigation of mobile robots is a new application field of control engineering. These subjects will get more and more important in the future not only because of the higher speed of the robots. Today the speed of mobile robots is approximately 0.5 m/s, in the future it will increase to 2.5 m/s and higher. Furthermore mobile robots are more and more confronted with fixed and moving obstacles which have to be included in the path planning algorithms. Fixed obstacles on the production level are e.g. machine tools, storage areas, walls, pillars, .. Moving obstacles are other robots, AGV`s as well as humans. There are a lot of very well known path planning algorithms for slow moving robots in an environment without obstacles available. In the framework of this research project three methods: path planning by means of fuzzy methods, by means of neural networks and as a combination of both path planning with neuro-fuzzy methods were adapted, implemented and tested. Most of the known methods are dealing only with theoretical subjects - experimental results are missing. One of the main goals of this project was to apply these 3 methods on commercial available mobile intelligent robot platforms. For the experiments a mobile robot platform (Nomad 200), a mini- robot (Kephera) and a mini-robot (Roby-Go) developed at the IHRT were available. The tests were carried out in standardized working environments. To minimize the time for these experiments simulation studies were carried out prior. The main result was that for practical applications all three tested methods are nearly equal. Fuzzy methods are partially successfully if the rule base is a balanced mixture between the number of the rules and the accuracy. Applying neural networks requires a time reduction for learning of the network. One possibility is to collect a sufficient number of pattern datas. This is also time consuming because the robot has to be operated manually and as much as possible data sets have to be stored. The practical experiments showed that dead zones appeared because of the angles of the sensors. Is a small obstacle in the dead zone the sensors are unable to recognize this object. Furthermore a disadvantage for the robot Nomad 200 was the arrangement of the sensors. Because of the difference in the height between the sensor ring and the floor flat obstacles were not recognized. In the future it would be necessary to improve the sensor technology dramatically. Meanwhile the results were successfully applied for the path planning and collision avoidance of fast moving mini- robots in working areas with up to 9 moving obstacles.