Sophisticated and Distributed: The Transportation Domain zy - zyxwvu Using DAI Methods in a Real-World Application*- K. Fischer N. Kuhn H.J. Muller+ J.P. Muller M. Pischel German Research Center for A.rtificia1 Intelligence (DFKI) Stuhlsatzenhausweg 3, DI-6600 Saarbriicken 11 Extended Abstract The transportation domain is presented zyxwvut as a multi- agent scenario, and the use of DAI techniques for solv- ing cooperatively the hard problems that occur within this domain is demonstrated. New analysis has shown that in Europe almost 40 zyxwvuts 96 of all trucks drive empty. Of course, this is an inac- ceptable state both for economical and ecological rea- sons. In fact, this could be avoided if more and better mechanisms of coordination and cooperation among the transportation companies were used. However, up tlo now there are no flexible systems for supporting distributed transportation scheduling currently. The transportation domain seems particularly well- suited for the use of DAI methods for the following reasons: zyxwvutsrqp 0 The domain is inherently distributed and decen- tralized. This holds for both knowledge and con- trol. The agents have local, possibly incomplete knowledge and local criteria for evaluation and decision-making . 0 Planning in the domain is highly dynamic, since new transportation orders may be entered into the system continuously, and since the actions of other agents have to be taken into consideration. 0 The shipping companies are autonomous agents. Task allocation, contracting, forming and exe- cuting joint plans! and negotiation are standard techniques applied within this domain for coor- dination among the agents. 0 The domain imposes some hard mathematical problems (e.g. the Scheduling Problems) which can be tackled by using knowledge-based, heuris- tic AI methods. In the following, we sketch the DAI techniques used in the MARS multi-agent scenario. MARS imple- ments a society of transportation companies whose goal it is to deliver a set of dynamically given orders satisfying given cost and/or time constraints. Hereby, three aspects are stressed: cooperation, task decompo- sition, and distributed scheduling. As regards coopera- tion, vertical (e.g. contracting between a company and its trucks and horizontal (e.g. between two compa- terns of cooperation such as mechanisms for combining intra- and interregional traffic and for exchanging or- ders to avoid rides without carriage are implemented. A negotiation model is provided to account for the decisions of the agents. Task decomposition 31 is es- sential both for allocating tasks to trucks and zyx c or offer- ing partial orders to other companies. The distributed scheduling problem is solved by giving the trucks local planning capabilities guided by local utilities. Trucks pass relevant informaition (e.g. about rides without carriages) to their company. The company uses this informaition to coordinate truck activities. The plan- ning and the execution phase in the MARS model are interleaved. Thus, the truck agents are able to in- tegrate incoming orders dynamically into their plans. This approach combines the advantages of both reac- tive and rational planning mechanisms. The MARS system has been implemented for a UNIX LAN environment using the multi-agent devel- opment language MAGSY[2]. First results are encour- aging. zyxwvu Six medium-sized companies (about 50 trucks each) run on six nodes of the network plan a load of 400 transportation orders’ between about 50 German cities in about 20 minutes and yield good solutions without using a central scheduling algorithm. For a more detailed description of the MARS system we re-. fer to [l]. References [l] M. Buchheit and N. Kuhn and J. P. Muller and: M. Pischel, “MARS: Modelling a Multiagent Sce- nario for Shipping Companies,” Proc. of the Eu-, ropean Simulation Symposium zyxwv (ESS 92), Dresden, 1992. [2] K. Fischer , “Concepts for Hierarchical Planning in a Flexible Manufacturing System,’’ PhD-Thesis, TU Miinchen, July 1992. [3] N. Kuhn and H.J. Muller and J.P. Muller. “Task Decompositions in Dynamic Agent Societies“ ,, to appear in: Proc. of the Intemational Symposium on Autonomous Decentralized Systems (ISA DS-93, Tokyo, Japan, 1993 nies) mec h anisms and protocols are investigated. Pat- ~~ ~ ‘This work has been supported by the German Federal Min- t e-mail: mueller@dlki.uni-sb.de . tion company for about three days! istry of Reseach and Technology under grant ITW 9104. ‘400 orders are enough to occupy a medium-sized transporta- zyx 454 1043-0989/93 $03.00 Q 1993 IEEE