The Integration of Explanation-Based Learning and Fuzzy Control in the Context of Software Assurance as Applied to Modular Avionics

Abstract

A Modular Power Management System (MPMS) is an energy management system intended for highly modular applications, able to adapt to changing hardware intelligently. There is a dearth in the literature on Integrated Modular Avionics (IMA), which has previously not addressed the implications for software operating within this architecture. Namely, the adaptation of control laws to changing hardware. This work proposes some approaches to address this issue. Control laws may require adaptation to overcome hardware degradation, or system upgrades. There is also a growing interest in the ability to change hardware configurations of UASs (Unmanned Aerial Systems) between missions, to better fit the characteristics of each one. Hardware changes in the aviation industry come with an additional caveat: in order for a software system to be used in aviation it must be certified as part of a platform. This certification process has no clear guidelines for adaptive systems. Adapting to a changing platform, as well as addressing the necessary certification effort, motivated the development of the MPMS. The aim of the work is twofold. Firstly, to modify existing control strategies for new hardware. This is achieved with generalisation and transfer earning. Secondly, to reduce the workload involved with maintaining a safety argument for an adaptive controller. Three areas of work are used to demonstrate the satisfaction of this aim. Explanation-Based Learning (EBL) is proposed for the derivation of new control laws. The EBL domain theory embodies general control strategies, which are specialised to form fuzzy rules. A
method for translating explanation structures into fuzzy rules is presented. The generation of specific rules, from a general control strategy, is one way to adapt to controlling a modular platform. A fuzzy controller executes the rules derived by EBL. This maintains fast rule execution as well as the separation of strategy and application. The ability of EBL to generate rules which are useful when executed by a fuzzy controller is demonstrated by an experiment. A domain theory is given to control throttle output, which is used to generate fuzzy rules. These rules have a positive impact on energy consumption in simulated flight. EBL is proposed, for rule derivation, because it focuses on generalisation. Generalisations can apply knowledge from one situation, or hardware, to another. This can be preferable to re-derivation of similar control laws. Furthermore, EBL can be augmented to include analogical reasoning when reaching an impasse. An algorithm which integrates analogy into EBL has been developed as part of this work. The inclusion of analogical reasoning facilitates transfer learning, which furthers the flexibility of the MPMS in adapting to new hardware. The adaptive capability of the MPMS is demonstrated by application to multiple simulated platforms.
EBL produces explanation structures. Augmenting these explanation structures with a safetyspecific domain theory can produce skeletal safety cases. A technique to achieve this has been developed. Example structures are generated for previously derived fuzzy rules. Generating safety cases from explanation structures can form the basis for an adaptive safety argument.