Genetic Fuzzy Tree AI toolkit for critical decisions

The Thales AI toolkit is a development and deployment suite that enables trustable, low power-consumption, embedded AI solutions by enabling interpretability and formal verifiability for complex autonomy, decision making, and predictive analytics for a variety of potential domains such as Aerospace, Defense, Banking, Medical, Automotive, among others.

The Thales AI toolkit is based on Genetic Fuzzy Trees (GFTs), a novel type of machine learning-based AI approach combining fuzzy logic, efficient network structures based on a human-like decision-making approach, and powerful Genetic Algorithms.

This innovative methodology was introduced by Psibernetix Inc. in 2015 and furthered by Thales since 2018. This GFT methodology fits the Thales “TrUE” AI approach; it is Transparent, Understandable, and Ethical.

At its core, the Thales AI toolkit is based upon fuzzy logic, developed by Lotfi Zadeh in 1965, which is a principal constituent of soft computing whose aim is to exploit the tolerance for imprecision and uncertainty to achieve robust andhigh performance control, at a low solution cost. Moreover, fuzzy logic, a "universal approximator", emulates human linguistic reasoning in a transparent and explainable manner to approximate, to any arbitrary degree of accuracy, any non-linear mapping between inputs and desired outputs in a computationally efficient manner.

This capability set led Fuzzy Logic to replace many standard mathematical/statistical systems due to its ability to reach extreme performance, while simultaneously being robust to randomness, uncertainty, and noise.

Thales’ research has greatly increased the scalability of fuzzy logic based AI systems such that the TrUE AI Toolkit can be utilized to create high performance, formally verified, and trustable AI systems for problems of high dimensionality in any of reinforcement, supervised, or unsupervised learning.

There is a need for “explainable AI” systems due the need to use AI in safety critical systems.  However, “black-box” style methodologies, such as Deep Learning Neural Networks (DLNN) are currently the principal AI approaches being currently investigated. There has been extensive work to make Neural Networks more explainable and trustworthy, with some progress being made. However, these primarily rely upon a second system to be made to attempt to explain the black-box AI.

Additionally, the black box AI research does not extend to interpretability, a key requirement to realizing the full potential of formal verification capabilities. Thales GFT technology is an AI/ML solution that can be inherently fully understood, trusted and formally verified.

Formal verification is the utilization of rigorous methods for the specification, design, and analysis of systems for verification of algorithms.  An example of the Thales AI Toolkits formal verification capabilities is the ability to mathematically guarantee adherence to safety specifications.

These specifications can take many forms such as design requirements, standard operating procedures, rules of engagement / international low, etc.  For example: “The AI must never ____” or “If state conditions are ____ than the AI output must always be within ____ or ____ ranges”.  Typically, these specifications are highly use-case specific, and there have been numerous aerospace and defense software errors that lead to disastrous results that could have been prevented by the ability to formally verify the system.

The formal methods tools utilized by Thales are incorporated into the GFT AI design and development process, and if cases are found that would lead to a failure to adhere to any set of the specifications, traces of input and intermediary values are provided.  Due to the interpretable nature of the Fuzzy Tree, these errors can be corrected during the training and development process cleanly, with holistic knowledge of how these edits to the AI to adhere to all requirements will affect the overall system.

The Thales AI Toolkit design process combines traditional numerical analysis for testing and evaluation with the formal verification capabilities to provide an unparalleled level of trustworthiness for a complex AI system.

For a discovery license, a development or a deployment license, please download our brochure and contact us for additional information.