Periodic Reporting for period 3 - WhiteMech (White-Box Self-Programming Mechanisms)
Periodo di rendicontazione: 2022-11-01 al 2024-04-30
White-Box Self-Programming Mechanisms
Context. We are witnessing an increasing availability of mechanisms that operate in nondeterministic (uncertain) environments and offer some form of programmability. These include manufacturing devices, smart objects and spaces, intelligent robots, dynamic business process management systems, and many others. All these mechanisms are currently being revolutionized by advancements in sensing (vision, language understanding) and actuation components (autonomous mobile manipulators, automated storage and retrieval systems). However, such mechanisms are held back by the fact that their logic is still based on hard-wired rules encoded in hand-crafted programs. WhiteMech aims at developing the science and the tools for a new generation of mechanisms to emerge: mechanisms that are able to program themselves, automatically tailor their behavior so as to achieve desired goals, maintain themselves within safe boundaries in a changing environment, and follow regulations and conventions that evolve over time. Crucially, empowering mechanisms with self-programming carries significant risks and therefore we must be able to balance power with safety. For this reason WhiteMech intends to realize mechanisms that are white-box, that is, whose behavior is at any moment fully analyzable and comprehensible in human terms, and guarded by human oversight. Remarkable recent discoveries by the applicant in Reasoning about Action and Generalized Planning in Artificial Intelligence, and their connections to Verification and Synthesis in Formal Methods, and Data-Aware Processes in Databases, chart an unanticipated novel path to produce a breakthrough in realizing powerful self-programming mechanisms, while keeping them human- comprehensible and safe by design. WhiteMech grounds its scientific results upon three driving applications: smart manufacturing (Industry 4.0) smart spaces (IoT) and business process management systems (BPM).
Main project objective: Studying self-programmable AI devices.
Manufacturing devices, smart objects, intelligent robots and other applications depend on mechanisms offering forms of programmability and relying on advanced sensing and actuation components. However, the effect of these mechanisms is limited because their logic relies on hard-wired rules encoded in hand-crafted programmes. The EU-funded WhiteMech project is developing the tools for a new generation of white-box self-programming mechanisms capable of automatically tailoring their behaviour to reach desired goals, support themselves with safe boundaries in a changing environment, and follow evolving regulations and conventions over time. WhiteMech bases its scientific results on combining knowledge and insight from autonomous sequential decision making in Artificial Intelligence, program verification and synthesis in formal methods, and data-aware processes in databases.
Context. We are witnessing an increasing availability of mechanisms that operate in nondeterministic (uncertain) environments and offer some form of programmability. These include manufacturing devices, smart objects and spaces, intelligent robots, dynamic business process management systems, and many others. All these mechanisms are currently being revolutionized by advancements in sensing (vision, language understanding) and actuation components (autonomous mobile manipulators, automated storage and retrieval systems). However, such mechanisms are held back by the fact that their logic is still based on hard-wired rules encoded in hand-crafted programs. WhiteMech aims at developing the science and the tools for a new generation of mechanisms to emerge: mechanisms that are able to program themselves, automatically tailor their behavior so as to achieve desired goals, maintain themselves within safe boundaries in a changing environment, and follow regulations and conventions that evolve over time. Crucially, empowering mechanisms with self-programming carries significant risks and therefore we must be able to balance power with safety. For this reason WhiteMech intends to realize mechanisms that are white-box, that is, whose behavior is at any moment fully analyzable and comprehensible in human terms, and guarded by human oversight. Remarkable recent discoveries by the applicant in Reasoning about Action and Generalized Planning in Artificial Intelligence, and their connections to Verification and Synthesis in Formal Methods, and Data-Aware Processes in Databases, chart an unanticipated novel path to produce a breakthrough in realizing powerful self-programming mechanisms, while keeping them human- comprehensible and safe by design. WhiteMech grounds its scientific results upon three driving applications: smart manufacturing (Industry 4.0) smart spaces (IoT) and business process management systems (BPM).
Main project objective: Studying self-programmable AI devices.
Manufacturing devices, smart objects, intelligent robots and other applications depend on mechanisms offering forms of programmability and relying on advanced sensing and actuation components. However, the effect of these mechanisms is limited because their logic relies on hard-wired rules encoded in hand-crafted programmes. The EU-funded WhiteMech project is developing the tools for a new generation of white-box self-programming mechanisms capable of automatically tailoring their behaviour to reach desired goals, support themselves with safe boundaries in a changing environment, and follow evolving regulations and conventions over time. WhiteMech bases its scientific results on combining knowledge and insight from autonomous sequential decision making in Artificial Intelligence, program verification and synthesis in formal methods, and data-aware processes in databases.
Self-programmability is essentially the ability of a mechanism to adapt to new, unexpected conditions, while still being able to achieve its goals. In this view, self-programmability requires that a mechanism be able to produce a plan or a strategy that achieves a desired goal, every time conditions change. In other words, the mechanism must be able to plan for a goal. Planning is a branch of AI that addresses the problem of generating a course of action to achieve a desired goal, given a description of the domain of interest and its initial state. The area is central to the development of intelligent agents and autonomous robots. One crucial element of Planning is having a model of the world that is used to perform synthesis for achieving the goal. In a sense, from the conceptual point of view, the essence of planning is program-synthesis under assumptions (assumptions being the model of the world). Planning is performed assuming that the world behaves in a certain way. However, in Planning typically the goal is simply reaching a desired state of affair. Instead WhiteMech want so state to consider agent task that can be very sophisticated. For this reason, in specifying tasks we adopt logical specification languages developed in Formal Methods for model checking (one of the most successful uses of logic in Computer Science). Specifically, we focus on Linear Temporal Logic (LTL), which is the specification formalism most used in Model Checking. However, when we use LTL to specify agent tasks, we adopt its finite trace variant, namely LTLf. The reason to focus on finite traces comes from the consideration that given a task, an intelligent agent should (1) reason on its model of the world, (2) synthesize a course of actions, i.e. the program to execute, (3) execute such program, and (4), when done, be ready for the next task. If the task requires an infinite execution, then the agent would reason only once in its lifetime and then execute the synthetized program forever. While this is perfectly fine if the reasoning is done by the designer as in Formal Methods, it is does not make much sense if the reasoning is done by the intelligent agent itself.
The research within WhiteMech has followed several paths, including, focussing on finite traces for specifying agent tasks, studying of Pure Past Temporal Logics, focussing on Non-Markovian Environment Specification, adopting multiple Environments Models and Synthesis of Best Effort strategies, investigation of Nondeterministic Strategies for resilience, developing a data-aware self-programming framework based on the Situation Calculus, studying of Read-Write Data Integration based on Ontologies, and merging Reasoning and Learning for Non-Markovian Reinforcement Learning and Restraining Bolts (the latter line of work is that it has deeply influenced the proposal of the ICT-48 Network TAILOR, which indeed includes a work-package (WP5) on “Deciding and Learning How to Act” that is based on these ideas). Moreover, selected science, tools, and techniques developed in WhiteMech are transferred to actual applications, ranging from Business Process Management (BPM) to Smart Manufacturing. Notably, the research done within WhiteMech has significantly influenced the “Augmented BPM Manifesto”, a novel manifesto on future forms of BPM.
The research within WhiteMech has followed several paths, including, focussing on finite traces for specifying agent tasks, studying of Pure Past Temporal Logics, focussing on Non-Markovian Environment Specification, adopting multiple Environments Models and Synthesis of Best Effort strategies, investigation of Nondeterministic Strategies for resilience, developing a data-aware self-programming framework based on the Situation Calculus, studying of Read-Write Data Integration based on Ontologies, and merging Reasoning and Learning for Non-Markovian Reinforcement Learning and Restraining Bolts (the latter line of work is that it has deeply influenced the proposal of the ICT-48 Network TAILOR, which indeed includes a work-package (WP5) on “Deciding and Learning How to Act” that is based on these ideas). Moreover, selected science, tools, and techniques developed in WhiteMech are transferred to actual applications, ranging from Business Process Management (BPM) to Smart Manufacturing. Notably, the research done within WhiteMech has significantly influenced the “Augmented BPM Manifesto”, a novel manifesto on future forms of BPM.
WhiteMech studies systems that operate in nondeterministic (partially known) environments and offer some form of programmability. These include manufacturing devices, smart spaces, intelligent robots, dynamic business process management systems, and many others. All these mechanisms are being revolutionized by advancements in sensing (vision, language understanding) and actuation components (autonomous mobile manipulators, automated storage and retrieval systems). However, such mechanisms are currently held back by the fact that their logic is still based on hard-wired rules encoded in hand-crafted programs. Instead, WhiteMech is demonstrating that it is possible to develop the science and the tools for systems that are able to program themselves automatically without direct human intervention. Naturally, empowering systems with self-programming may carry significant risks. For this reason it is important to develop systems, whose behavior is analyzable, comprehensible and explainable in human terms, and guarded by human oversight. Therefore, WhiteMech advocates developing synergies between AI and several other fields of CS, most prominently Formal Methods.