| Indoor wireless positioning technology has significant application prospects in civil and military fields.Compared with outdoor location technology,indoor wireless positioning technology has the characteristics of complex and variable environment,serious multi-path propagation,and high accuracy requirements.At present,various research organizations and individuals have proposed various indoor positioning systems for localization signals of different devices,such as UWB,Wi Fi,geomagnetic,Bluetooth,infrared,etc.Traditional RSS fingerprint-based indoor localization methods require establishing a fingerprint database in the offline phase and using the fingerprint data in the offline phase to train a model for the online phase,and the RSS signal distribution in the online phase needs to be similar to the RSS distribution in the offline database building phase.However,in the actual indoor positioning application scenario,there are various problems with this fingerprint database positioning method:(1)Most indoor positioning techniques obtain a single type of signal source,and the heterogeneity of signal sources,differences in equipment models,and changes in signal strength can have an impact on the positioning accuracy;(2)The robustness of fingerprint database location;over time,changes in the temperature and humidity of the location environment,adjustments in the layout,or increases or decreases in the number of access points can cause changes in the RSS distribution.(3)The establishment of an offline fingerprint database is timeconsuming and labor-intensive.The distributed exploration,cooperative learning,diverse experience summarization,and robustness of Multi-Agent Reinforcement Learning have natural advantages for solving the above points.In this thesis,two indoor multi agent localization algorithms based on reinforcement learning are proposed to solve the above problems faced by traditional RSS fingerprint localization,and experiments are conducted on actual data to verify the effectiveness of these two algorithms.The primary focus of this thesis is as follows.In this thesis,we improve the Dual Depth Q Network and apply it to Multi-Agent Systems.We propose a multi-agent indoor location algorithm based on the Dual Depth Q Network.This algorithm uses a small amount of location tag data at the initial stage for unsupervised learning,avoiding the cumbersome process of establishing a fingerprint database,and adapting it to indoor location scenarios by setting reasonable reward functions.Experimental results on measured data demonstrate that the proposed algorithm can alleviate the problem of decreased positioning accuracy caused by differences in the distribution of RSS data to a certain extent,and has higher positioning accuracy than other commonly used algorithms.This thesis proposes a Multi-Agent indoor location method based on Evolutionary Transfer Reinforcement Learning Framework,which updates the network parameters of the target agent through learning the behavior of the source agent to dynamically adapt to changes in the indoor environment,to overcome the problem of RSS data distribution differences,and still has good adaptability to the indoor dynamic environment without establishing a fingerprint database,By comparing this method with other commonly used positioning methods through open data sets,simulation experiments in various aspects have verified the effectiveness and practicality of the method proposed in this article. |
