LEARNING DYNAMIC PROPERTIES OF AN ENVIRONMENT USING MULTIPLE AGENTS

Abstract

If we place an agent in an environment and leave it to the agents to learn the map of the envi- ronment, then there are many ways to learn the static topological/metric map of the environment, the problem considered in this thesis is to incorporate the discovering of timely patterns of changes getting introduced in the map of the environment. But the environment keeps changing in the real world, for example, in an organization the gates closes and opens at particular timings , resulting in path obstruction and path opening and if we can learn those patterns of path getting obstructed and opening , then we can incorporate those learning which can help the agent to decide (at the starting only) upon the most pro table path to follow instead of discovering blocked paths later. The challenging aspect here involves two stages, rst is to detect the states of the paths as either open/closed and second is to learn these time patterns of when do a path gets blocked/reopened. The given task is very di cult and time-taking for a single-agent system but if multiple-agents disseminate in the environment and try to learn the time-patterns of path obstruction in di erent parts of the environment, then they can at later stage combine this information collected by every agent. This way the learning process can be increased exponentially Along with solving above given problem, we also try to look at many other sub-problems of collision avoidance during path propagation and noise removal from sensor readings which are ultimately going to be a part of the solution.