Using mainstream AI/ML algorithms, entire classes of practical use cases are doomed to be computationally intractable or require many multiple universe-lifetimes of data collection. I/ITSEC paper 19149 (Allen) introduced an innovative architectural approach to AI where neural sigmoid functions are replaced by mathematical models of arbitrary complexity, thus collapsing net sizes and depth, and ultimately reducing computational and data requirements. Without the constraints imposed by neural assumptions, mathematical models may be nonlinear and/or discontinuous and may be guided by human knowledge of the system. This novel approach, along with advanced optimization methods presented in I/ITSEC paper 19109 (Allen), forms the basis a new family of Evolved AI solutions.
During a National Defense (January 2020) interview, NDIA’s Senior Fellow for AI expressed how algorithms and framework have evolved beyond supervised learning into unsupervised and reinforcement learning. Having presented the algorithms and laid the framework for Evolved AI, the focus of this paper shifts to applications of this emerging technology to stochastic clustering.
The paper first describes how the laptop-executable approach combines elements of both hard and soft clustering without the need for cleaning/scaling data, nor the need for training data. Unlike k-means and k-medoids, the cluster number (k) is not needed a priori, as with hierarchical clustering. By leveraging fuzzy c-means and Gaussian mixture models, data points may belong to more than one cluster having different sizes and correlations. Overall, cluster number is adaptively determined from the distribution of resultant cluster permutations.
The paper then presents and discusses an example in the context of pulse spectrum analysis where preliminary work in applying multi-dimensional stochastic clustering has proven successful.
Upon summarizing results, the paper concludes by recommending applications of this emerging technology to Training and Education, for example measuring pilot training exercise data and clustering results in terms of ideal execution.