Implementing Augmented Intelligence In Systems Engineering

This paper will explore the opportunities for artificial intelligence (AI) in the system engineering domain, particularly in ways that unite the unique capabilities of the systems engineer with the AI. This collaboration of human and machine intelligence is known as Augmented Intelligence (AuI). There is little doubt that systems engineering productivity could be improved with effective utilization of well-established AI techniques, such as machine learning, natural language processing, and statistical models. However, human engineers excel at many tasks that remain difficult for AIs, such as visual interpretation, abstract pattern matching, and drawing broad inferences based on experience. Combining the best of AI and human capabilities, along with effective human/machine interactions and data visualization, offers the potential for orders-of-magnitude improvements in the speed and quality of delivered. INTRODUCTION Augmented Intelligence (AuI), an approach that promotes “team play” of human and machine intelligence, is a modern refinement of established AI approaches. By effectively joining the human skills in pattern matching, unstructured data, and intuition with computational approaches that excel in domain search, systematic trade space exploration, and statistical evaluation, the combined “team” has been proven to be more effective than either in isolation. For instance, machine learning algorithms can process past system designs, learn significant design characteristics, and visually present outcomes and the various tradeoffs. The human team can evaluate the domain space quickly, and watch for exceptional cases that might not be accurately handled by the machine. This paper will explore the potential, challenges, and requirements of implementing AuI in the engineering of systems. AuI has been enabled by the adoption of Model Based Systems Engineering (MBSE), and particularly the use of formal modeling languages such as the Systems Modeling Language (SysML), UML, Architecture Analysis and Design Language (AADL), etc. Previous document-centric approaches to systems engineering resulted in less well-defined systems that, while generally intelligible to human readers, were too unstructured for algorithmic approaches. The model-centric approach, using SysML, is ideal for AuI, since SysML was designed to be both human and machine readable. For the human, there is a concrete visual Proceedings of the 2018 Ground Vehicle Systems Engineering and Technology Symposium (GVSETS) IMPLEMENTING AUGMENTED INTELLIGENCE IN SYSTEMS ENGINEERING Page 2 of 15 representation easily interpreted by a human (e.g. the familiar SysML block diagram, containing “boxes and arrows”). For the machine, SysML has a precise semantic representation that allows for simulation and model execution. Together, these two representations make SysML well-suited as a language for enabling AuI. The innovation of systems is a prime opportunity for the application AuI, given the rapid increase in product complexity and time constraints. As an example, the relationship between the requirements of a system and the functional performance (e.g. SWAP-C: space weight, power, and cooling) is fundamentally mapping, traceability, and parametric relationships between requirements and design parameters. In other words, if a requirement is changed, what is the impact on functionality and performance? Even with MBSE, this is still a manual process of evaluating the models, collecting the inputs of domain experts, and determining the impact. In contrast, AuI offers the ability to use every past system as inputs to machine learning algorithms. The design team can quickly visualize the opportunities and the challenges of those design decisions. Not just one course of action can be evaluated, but every course of action could be visualized, evaluated, and communicated to the design team, subject more to computational constraints rather than time constraints. MODEL-BASED SYSTEMS ENGINEERING (MBSE) The growing complexity of systems necessitates a systems engineering approach. It requires a systems paradigm which is interdisciplinary, leverages principals common to all complex systems, and applies the requisite physics-based and mathematical models to represent them. INCOSE defines Model-Based Systems Engineering (MBSE) as “the formalized application of modeling to support system requirements, design, analysis, verification and validation activities beginning in the conceptual design phase and continuing throughout development and later life cycle phases...” [1] The Object Management Group’s MBSE wiki notes that “Modeling has always been an important part of systems engineering to support functional, performance, and other types of engineering analysis.”[2] The application of MBSE has increased dramatically in recent years and is becoming a standard practice. This has been enabled by the continued maturity of modeling languages such as SysML and significant advancements made by tool vendors. These advancements are improving communications and providing a foundation to integrate diverse models. MBSE is often discussed as being composed of three fundamental elements – tool, language and method. The third element, method, has not always been given proper consideration. Because the language and tool are relatively method independent, it is methodology which further differentiates the effectiveness of any MBSE approach and its ability to help manage the complex and interrelated functionality of today’s systems. For the approach discussed in this paper, the “methodology” includes the application of Artificial Intelligence to augments the application of Systems Engineering activities. ARTIFICIAL INTELLIGENCE Artificial Intelligence (AI) is the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. Since the term AI is widely used for a variety of different algorithms and approaches, most definitions are functional rather than technical. For instance, according to Stanford AI researcher Jon McCarthy, “Any Proceedings of the 2018 Ground Vehicle Systems Engineering and Technology Symposium (GVSETS) IMPLEMENTING AUGMENTED INTELLIGENCE IN SYSTEMS ENGINEERING Page 3 of 15 program can be considered AI if it does something that we would normally think of as intelligent in humans” [3]. There are two primary categories of AI: Rules Based: Rules based AI uses a set of defined rules to derive and manipulate data. This represents explicit knowledge that can be provided to the AI system. For instance, in lexical parsing of the English language for natural language processing, grammatical rules can be defined (e.g. proper nouns are capitalized, as is the first word in a sentence). Rules based is also used for solvers, such as navigation algorithms that minimize travel time while following road rules (e.g. turns are permitted at certain intersections). Patterns Based: Usually known as Machine Learning (ML), patterns based approaches seek to capture tacit knowledge knowledge which is difficult or impractical to explicitly define through statistical approaches. For instance, it is difficult to completely list rules for email spam filtering. However, with a set of emails categorized as either spam and not spam, an algorithm can infer certain words or phrases that are effective predictors of spam for classification of new emails. Similarly, autonomous systems can learn appropriate driving techniques by observing a human driver. Many modern systems are hybrids of the rules and pattern approaches. For instance, chess AI systems use rules (e.g. 6 piece types, each with a small set of moves) for traversing the game tree, along with learned patterns to preemptively eliminate certain branches as poor moves. Proceedings of the 2018 Ground Vehicle Systems Engineering and Technology Symposium (GVSETS) IMPLEMENTING AUGMENTED INTELLIGENCE IN SYSTEMS ENGINEERING Page 4 of 15 HUMAN / AI INTERACTIONS Competition Chess was one of the first significant applications of AI. It was an ideal game to test the capabilities of AI, since it is a zero-sum, perfect information two-person game with a small set of rules. Zero-sum refers to the fact that an advantageous move by player A is disadvantageous to player B. Perfect information means that all positions for each player are perfectly visible (in contrast to other games like poker, where some information is hidden, and there is a role for “bluffing”). Chess is also a Markov process, where each arrangement of chess pieces can be evaluated independently of the moves that created that arrangement. These characteristics allowed for effective implementations of state space search algorithms, and comparisons of performance of human player versus machine [4]. Figure 1 shows the evolution of chess AIs as ranked on the ELO scale, a method for calculating the relative skill levels of players in zero-sum games such as chess (named after its creator, physicist Arpad Elo) from 1984 to 2016. On the second Y axis is MIPS (millions of instructions per second) per $100, the affordability of computational power shown in logarithmic units. As computational power grows exponentially, the chess Figure 1 Growth of chess AI capabilities as measured on the ELO ranking scale of zero-sum games. Human grandmaster performance is shown in dashed-red. The line is approximately horizontal, as human performance changes very little on the scale required to show AI growth. Proceedings of the 2018 Ground Vehicle Systems Engineering and Technology Symposium (GVSETS) IMPLEMENTING AUGMENTED INTELLIGENCE IN SYSTEMS ENGINEERING Page 5 of 15 AI performance grew linearly, characteristic of tree search algorithms. Shown in red is the best human grandmaster performance, which changes very little during this timeframe. In 1996, IBM’s supercomputer Deep Blue pla