Resource Page: Research Project Details
Learning in Multi Agent System for Multi-actor spatial planning
Institution: Centre for GeoInformation Sciences
Principal Investigators: Bidhyananda Yadav
Description: Although the GIS has been realized as an important analytical tool for solving spatial explicit problems, it fails to give desired result when situation is complex, particularly when it comes to represent human behavior and their intensions. Actually, the interaction between the environment and the stakeholders is inherently quite complex. Mathematical modeling tools are not able to support the soft knowledge used by humans in decision making and hence GIS cannot support it either. Thus, newer insights are necessary to understand and model such complex and dynamic interactions. Recent developments in the field of artificial intelligence and particularly in that of multi agent system (MAS), looks promising to solve these complex dynamic problems. Coupling GIS with MAS seem an option to represent the soft knowledge used by humans in problem solving. With such a system it should also be possible to model and simulate the complex dynamics of spatial interactions.
Spatial planning nowadays is an extremely sensitive issue in many parts of the world involving multiple actors/stakeholders. Moreover, many complex political and social processes influence it. So, it is a highly decentralized process (at least in the democratic societies) and the process can last for years. The final outcome of the process is not always desirable for the all the stakeholders and hence conflict arises. This is one of the reasons why people use models to study the effects of land use planning. However, these models are far from representing the reality, and have to make many assumptions. So the final outcomes of these models are not correct as well.
One of the main drawbacks of these models is that they fail to capture the desires and the preferences of the actors, major driving force for most land use change, involved in the planning process. To address this shortcoming, MAS models have gained popularity and scientists are increasingly applying these models to study land use change for multi-actor planning. While it is possible to capture the human desires and preferences through these models, there are still many issues that need to be addressed at large. For instance, how to validate such models, how to initiate and handle negotiation between agents and how each agent can learn about others so as to come with better decisions in the end.
Multi-actor spatial planning involves the interaction among various actors and lots of negotiation and communication is going on in between them. In this process the actors also try to learn about others desires and preferences so that they can come at consensus and the final result is acceptable to a majority of actors. Since, MAS seem to be a feasible alternative to design such systems, learning among agents in a multi agent framework is an important issue to be tacked at large.
One of the reasons why the systems designed on multi agent framework for spatial planning lacks credibility is that the agents are annoying stupid which is in sharp contrast to what agents should be. To overcome this situation, there is a strong need to equip agents with some learning capability. While still there is no consensus among the research community on what is intelligence and what is learning, there is a strong agreement that there can be no intelligence if the system does not have the ability to learn.
Equipping agents with the ability to learn mean that intelligent systems can be developed which can have some degree is fault tolerance and can act even in the unpredictable and dynamic environment, which is actually the case with earths systems and processes.
Having said all this I want to stress that the general purpose of this research work is to investigate about applying various learning techniques that can be used for multi-actor land use planning in a multi agent framework. There is a need for a kind of artificial environment in which policy can be developed and tested in order to cope with the increasing complexity of reality. Hence, the broader notion of this research is to aid in the developed of such agent based models where policies and plans can be evaluated and their effects appropriately studied before they are put into practice.
Software Used: Java, Repast and ArcView
Contact: Bidhyananda Yadav