A learning algorithm of dynamical associational multi-agents for intelligent environments

An intelligent inhabited environment applying interconnected embedded agents by network has intelligent reasoning, planning learning, and control capabilities. Thermal and light comforts are two major control objectives for the environment to deal with using data-driven control method. Practically, dynamic association level of agents should be learned from online data with three reasons: changing structure of agents with the devices to be added to or removed from the environment during residents' life, a large number of dimension of input and output vectors making it is very difficult to design learning based controller, and a multitude of interconnected embedded agents resulting in major load in network communication and calculation. This paper presented a novel online learning algorithm to obtain the structure agents with different functions through identifying the associations between inputs and outputs of the environment. An association weight matrix can be calculated online and the embedded agents can be dynamically divided into multiple subgroups. This can reduce dimension of input vector for each subgroup, reducing network communication load among embedded agents, decreasing the complexities of programming, and improving the learning rate of agents. The experiment results demonstrated the effectiveness and significance of the learning algorithm.