Self-* Programming: Run-Time Parallel Control Search for Reflection Box

Abstract

Real life situations may require an automatic fast update of the control of a plant, whether the plant is an airplane that needs to overcome an emergency situation due to drastic environment change, or a process that needs to continue executing an application in spite of a change in an operating system behavior. Settings for run-time control synthesis are defined, assuming the environment maybe totally dynamic, but is reentrant and history oblivious for long enough periods. A reentrant environment allows several copies of a plant to interact with the environment independently; a history oblivious environment ensures a repetition (in the probabilistic case with the same probability) of an interaction starting with a plant in a certain state and replaying its output to the environment. Total dynamic changes of the environment do not allow a definition of weakly realizable specifications, as weakly realizable specifications depend on the environment behavior. On line experiments of the environment assists in the implementation of unrealizable specifications; Automatically checking whether the unrealizable specifications define weakly realizable specifications, given the behavior restriction of the current environment. A successful search for a control implicitly identifies the weakly realizable specifications, and explicitly the implementation that respect the specifications. Different settings and capabilities of the plant are investigated. In particular, (i) plant state reflection that allows observation of the current state of the plant, (ii) plant state set that generalizes the reset capability, allowing setting the plant to each of its states, and (iii) (static or dynamic) plant replication that allows instantiation of plant replicas or use of preexisting plant replicas for parallelizing testing algorithms. The algorithms presented prove that the above capabilities enable a polynomial search for a new control upon a drastic change of the environment.

Labor Division with Movable Walls: Composing Executable Specifications with Machine Learning and Search (Blue Sky Idea)

“…These are manifested in a variety of forms, in development tools, testing, risk analysis, and of course in the running system itself. Specific examples include applying rigorous software engineering methods in developing collaborating agents (Jiménez, Piattini, and Vizcaíno 2009;Sturm and Shehory 2014), recent efforts to apply formal verification techniques to neural nets (Katz et al 2017), using AI algorithms for speeding up automated program synthesis and formal verification in bounded model-checking (Biere et al 2009), and using AI planning algorithms for run-time lookahead in formally-specified systems (Harel and Segall 2007;Weinstock 2015;Brukman et al 2015). Additional examples can be found in (WISE Intl.…”