A Fuzzy-Based Machine Learning Model for Robot Prediction of Link Quality
With foresight into the state of the wireless channel, a robot can make various optimization decisions with regards to routing packets, planning mobility paths, or switching between diverse radios. However, the process of predicting link quality (LQ) is nontrivial due to the streaming and dynamic nature of radio wave propagation, which is complicated by robot mobility. Due to robot movement, the wireless propagation environment can change considerably in terms of distance, obstacles, noise, and interference. Therefore, LQ must be learned and regularly updated while the robot is online. However, the existing fuzzy-based models for assessing LQ are non-adaptable due to the absence of any learning mechanism. To address this issue, we introduce a fuzzy-based prediction model designed for the efficient online and incremental learning of LQ. The unique approach uses fuzzy logic to infer LQ based on the collective output from a series of offset classifiers and their posterior probabilities. In essence, the proposed model leverages machine learning for extracting the underlying functional relationship between the input and output variables, but deeper inferences are made from the output of the learning algorithms using fuzzy logic. Wireless link data from a real-world robot network was used to compare the model with the traditional linear regression approach. The results show statistically significant improvements in three out of the six real-world indoor and outdoor environments where the robot operated. Additionally, the novel approach offers a number of other benefits, including the flexibility to use fuzzy logic for model tuning, as well as the ability to make implementation efficiencies in terms of parallelization and the conservation of labeling resources.
USMA Center/Institute Affiliation
Robotics Research Center
C. J. Lowrance, A. P. Lauf and M. Kantardzic, "A fuzzy-based machine learning model for robot prediction of link quality," 2016 IEEE Symposium Series on Computational Intelligence (SSCI), Athens, 2016, pp. 1-8. doi: 10.1109/SSCI.2016.7849899