Application of the Hierarchical Temporal Memory to Smart Transport

Traffic congestion has a significant effect on the economy and environment. There is a need for more innovative solutions that adopt new technologies to tackle the congestion issue to reduce journey times and fuel emissions. Recently, there have been many attempts to address this issue using different machine learning approaches such as the k-nearest neighbour algorithm and artificial neural network. However, additional measures and new methods need to be taken and investigated respectively to resolve the traffic congestion issue.

Due to new advanced IoT technologies, the data generated by sensors are huge and continually changing. Thus, there is a need for a machine learning algorithm that is capable of dealing with this huge amount of data. One of the new advancements in machine learning models is the Hierarchical Temporal Memory (HTM), a neuroscience-based algorithm that attempts to mimic the neocortex functions in the human brain. This made HTM a good choice since the memory requirements for the HTM algorithm are less than compared to machine learning algorithms.

To the author’s knowledge, using HTM for the smart transport application is not commonly adopted in the literature, and more research is needed. Therefore, a proposal of an HTM theory-based framework is presented in this work. The novelty of this framework is the use of multi-level anomaly detection with a predictive feedback network. Each level comprises different inputs that are combined to an upper level to detect abnormalities from previous patterns in the lower levels. It then introduces a predictive level model to provide feedback by using the patterns learnt from the combined upper levels of the HTM hierarchy.

To evaluate and test the proposed framework, several synthetic and real data sets were reviewed. Based on the results of the review, a pre-processing part of the Highways England’s Motorway Incident Detection and Automatic Signalling (MIDAS) traffic database was required to evaluate the performance of the proposed framework.

The framework was implemented using the latest Numenta Platform for Intelligent Computing (NuPIC), which is an implementation of HTM. The framework was evaluated using two sets of input data from the pre-processed MIDAS datasets. The parameters of the encoders, spatial pooler and temporal memory were optimised to produce the highest accuracy and F-measure. The results show a rise in both accuracy and F-measure by 2% and 6.5% respectively.

The proposed framework performance is evaluated against a number of conventional anomaly detection methods using F-measure as one of the performance metrics. The analysis of the results revealed that the proposed framework achieved a 60% F- measure, which outperforms the K-Nearest Neighbour Global Anomaly Score (KNN-GAS) by 35%, the Independent Component Analysis-Local Outlier Probability (ICALoOP) by 34% and the Singular Value Decomposition Influence Outlier (SVD-IO) by 33%. These results provide evidence that the proposed HTM based framework can achieve better performance when dealing with traffic detection datasets.