This research presents an innovative approach to help engineers with the design of a complex engineered system. The design of complex engineered systems undertaken with respect to a variety of design objectives is substantially comparable to the multi-agent coordination problem. In both cases, the decisions at the component level subsystems and agents , and the interaction between those components, lead to global behavior. In multi-agent coordination, a key research challenge is to determine what each agent needs to do so that the system as a whole achieves a predetermined objective.
Figure 1 illustrates the design process envisioned in this paper. The approach we explore is to implement a team of autonomous agents responsible for selecting the best concept using multi-agent coordination. After the customer and engineering requirements are defined, engineers will create a team of agents suitable for the problem related to the system-level objectives.
A cooperative coevolutionary algorithm CCEA will perform the design exploration and multi-agent coordination. The algorithm will autonomously evaluate, select, and refine the design solution that results from the best trade-offs between all the subsystems. The overall goal of this research is to develop a design methodology that will help engineers to design complex engineered systems using a multi-agent coordination approach.
To be able to achieve more complex solutions, the autonomous agents will coordinate their actions design decisions to optimize the multi-objective system. Sections 2. Also included is a discussion on integrated concurrent engineering ICE to design complex engineered systems. The concepts of multi-objective optimization and multidisciplinary design optimization MDO , which are important methods and concepts to this research, are also presented within this section.
Selection of design architecture while considering various design criteria and sources of uncertainty is a fundamental research problem in designing complex systems. Explicitly computing quantitative and qualitative objectives of a complex system is viewed as the preferred method for formalizing the design process; however, one of the fundamental problems in typical large-scale engineering system design is the overemphasis on requirement satisfaction for evaluating design alternatives [ 1 ]. This focus is primarily the result of the acquisition process but is exacerbated by overly simplistic design objectives, such as minimizing weight or cost, which do not reflect the actual value of the designed system.
As an example, rather than making design decisions based primarily upon requirement i. This is a dramatic change in perspective for system design, promising a reduction or elimination of cost and schedule overruns [ 3 , 4 ] by identifying high-value designs for development. Value-centric design can be considered part of the larger field of decision-based design DBD [ 5 , 6 ].
DBD has been specifically developed in the system design community as a decision-theoretic approach to selecting a preferred system design from among the alternatives. DBD takes an enterprise-level view of the design problem, considering not only typical engineering concerns but also broader objectives that comprise the total value of the system to the enterprise.
Model-based design techniques are often needed when designing complex engineered systems [ 7 ]. The model-based design approach builds a model of the system that is used to simulate the functions and evaluate the performance. These models are at times segmented and combined to form a single model, although they do not need to be [ 7 ]. The simulation of the models helps designers to understand the behavior and performance of the system before any physical testing.
Most of the times, models such as nuclear power plants are so large and complex that they are too computationally expensive to create, let alone simulate. For this type of problems, the system is divided into smaller subsystems that can be simulated separately to collect information. However, this approach outcome with high levels of uncertainty because of the missing information of the complete model simulation.
Complex engineered systems present emergent interactions that are not necessarily modeled or expected couple a set of components or subsystems together. The degree of this coupling , or complexity as it is sometimes called, is difficult to determine because of the associated ambiguity related to the unknown component links. Coupling is described by the relationship between system variables and functions [ 8 ]. Nevertheless, it could be possible to shift the focus from modeling interactions to evaluate and incentivize subsystems so that their collective behavior achieves the system design goals.
This shift in focus could be translated into a new methodology to model complex engineered systems. A multi-agent coordination approach could be used to determine how to distribute responsibilities in a design process to the components in the design that are crucial to the success of the system. Integrated concurrent engineering represents a collection of practices that attempt to eradicate inefficiencies in conceptual design and reorganize the process of sharing information inside a design team.
ICE uses a combination of experienced designers; state-of-the-art modeling, visualization and analysis tools; social processes, and a specialized design facility; to conceive preliminary designs for complex engineered systems.
When compared with a traditional sequential engineering method, ICE users cut project schedule by several orders of magnitude, while considerably improving design cost and maintaining quality standards. Within a brief time, ICE allows engineers to consider, implement, evaluate, accept, and reject various ideas, with a relatively high level of fidelity [ 9 ].
Traditional design approaches constrain interdisciplinary trades because of a deficiency of communication among team members. Information is often dispersed throughout the project team, meaning those searching data on a particular subject have no central location to search [ 10 ].
As a result, engineers spend a significant amount of time searching or recreating information that already exists, rather than developing new information or data. The ICE structure allows teams to work independently on problems local to a specific subsystem and to coordinate effectively on problems that affect other teams. Projects using ICE methodology are more flexible and can quickly adapt to changes in top-level requirements. As a result of the combination of all these factors, engineering teams that work with ICE can complete rapid trades among complex multidisciplinary subsystems.
Whereas traditional engineering superficially resembles a government bureaucracy, ICE performs the same work in an environment similar to NASA's shuttle mission control operations. In system design, multiple criteria, such as cost, safety, and performance, are considered in the process. This represents the multi-objective design optimization MOO problem [ 15 — 18 ]. General methodologies exist for solving MOO problems [ 19 , 20 ]; the distinguishing feature of a MOO problem is that, in general, one cannot identify a single solution that simultaneously optimizes each objective.
In the design objective space, the point obtained by solving the k single-criteria optimization problems individually is called the utopia ideal point. The utopia point is not achievable due to conflicts among the multiple design criteria. The Pareto frontier consists of Pareto solutions in the objective space such that a design criterion cannot be further improved without sacrifice to another criterion. Various methods for finding the Pareto frontier have been developed, such as the weighted-sum method [ 22 ], compromise programming [ 15 ], and the normal boundary intersection method [ 23 ].
Work exists in identifying the Pareto frontier when uncertainty exists in the multiple objectives, either by treating the mean and variance of each objective as separate objectives [ 24 ] or by computing an expected single attribute utility for each objective [ 25 ]. In this research, the ultimate goal is to optimize a set of multiple objectives, which is one of the distinctive characteristics of a complex system.
This research will enable the identification of the Pareto frontier and allow significant trade-offs in problems with nonconvex objective functions, utilizing agent-level information. Multidisciplinary design optimization is an optimization method used in engineering problems that involve multiple subsystems [ 10 ]. The advances in computing power have increased the popularity of this type of optimization methodologies. Some techniques have emerged in an attempt to integrate both system decomposition and optimization such as collaborative optimization CO and bi-level integrated system synthesis BLISS.
Collaborative optimization is a multidisciplinary design optimization technique, developed at Stanford University, that divides a problem along disciplinary lines into subproblems. Each subproblem is then optimized so that the difference between the obtainable subsystem performance and specific variables chosen by the system optimizer is minimized [ 26 , 27 ]. This methodology is powerful and efficient for problems with well-defined disciplinary boundaries, a large number of shared variables and calculations, and a minimum of interdisciplinary coupling. However, on the downside, CO leads to setups with high dimensionality, which requires high processing power.
CO has been successfully used in several engineering problems typically in the area of vehicle design. Bi-level integrated system synthesis is a multidisciplinary design optimization technique, developed at NASA Langley Research Center [ 28 — 30 ], that uses hierarchical decomposition. BLISS works similar to CO, where the goal is to optimize distributed engineering systems created by experienced groups who work concurrently to solve a design problem.
The methodology differs from CO because preference weights are used for multi-objective optimization at the subsystem level.
The constraints and coupling variables are also controlled fairly differently. In the first group, the variables are optimized at the local level and are found only within each of the subsystems. The second group contains variables, which are outputs by one subsystem and are used as inputs for a different subsystem. Finally, the third group includes the system-level design variables, which are shared by at least two subsystems. One of the biggest problems in modern design results from the conflict between the problem decomposition and MDO [ 10 ]. Decomposing a problem into smaller subproblems makes the overall problem more controllable, but as a result, it is more challenging for system-level optimization to make a significant contribution.
Connecting a different number of subsystems into one is not simple. As the complexity of the system increases, so does the complexity of the model needed to complete the system-level optimization. Solving optimization problems with a computer can take extended periods, which can result in a delay of time between members on a design team. While waiting for the optimization results to return, the design team continues with their work, often updating models and reacting to changes in the requirements.
When an optimization does finally produce data, the results are often obsolete by these changes. This is the main weakness of MDO because it prohibits a full integration of the parts, subsystems, and teams in the design. Thus, it is necessary to have a methodology that can relieve the fundamental conflict between these two approaches throughout the design cycle. Engineers invest a lot of effort to create accurate models for complex engineered systems. The desired objective is to use a model that allows engineers to perform accurate failure analysis and evaluations of the system before any manufacturing process.
The modeling of complex engineered systems and the assessment of failures are highly researched areas of interest within the engineering literature. This includes different areas that cover topics from performance evaluations to sensitivity analyses with input variables under uncertainty. The primary objective of this research is to develop a methodology where a team of engineers can design complex engineered systems with the implementation of a multi-agent coordination approach.
The design of complex engineered system undertaken with respect to a variety of design objectives is fundamentally similar to the multi-agent coordination problem. In both cases, the decisions at the component level subsystems or agents and the interactions among those components lead to global behavior complex system or multi-agent system. Sections 3.
Multi-agent coordination is a key research area in agent-based approaches to automation [ 31 ]. One of the biggest challenges in such an approach is decentralization of control and, in particular, the question of how to incentivize the individual agents such that they work together [ 32 ] to achieve the system objective.
The key challenge is that a system designer needs to address two major credit assignment problems: structural and temporal [ 31 , 32 ] credit. The first addresses who should get credit or blame for system performance, and the second addresses which key action at which key time-step is responsible for fulfilling the objective [ 33 , 34 ]. The temporal credit assignment problem has been extensively studied through single-agent reinforcement learning [ 32 , 35 ]. The structural credit assignment problem has also received attention, and has been addressed by two broad approaches: feedback shaping and organizational structures.
Feedback shaping aims to shape the system objective such that the action of agents optimizing local objectives results in desirable system-level performance [ 36 , 37 ].
Organizational structures decompose the agents themselves into roles that enable coordinated behavior [ 38 , 39 ]. One particular research area in the credit assignment problem focuses on ensuring that agents' objectives are aligned with the system objective i. Providing agents with objectives that satisfy these two properties formalized in Refs. A particular set of agent objectives that achieve these goals is the difference objectives , which is based on the difference between the actual performance of the system and the performance of a counter-factual system in which certain agents have been removed.
Difference objectives have been extensively studied and applied to real-world applications including air traffic control, multirobot coordination, and resource allocation [ 40 , 42 — 44 ]. The success of the difference objective approach in developing appropriate agent learning objectives suggests that the method applies to complex system design where a structural credit assignment problem exists when designing individual components.
One implementation of this approach, based on coevolutionary algorithms, is described next. Evolutionary algorithms are a class of stochastic population-based search algorithms, which can often outperform classical optimization techniques, particularly in complex domains where gradient information is not available [ 45 ]. An evolutionary algorithm typically contains three basic mechanisms: solution generation, a mutation operator, and a selection operator.
These mechanisms are used on an initial set of candidate solutions, to generate new solutions and retain solutions that show improvement. Simple evolutionary algorithms are excellent tools but need to be adjusted to apply to large multi-agent search problems for distributed optimization. One such modification is coevolution , where multiple populations evolve simultaneously to develop policies for interacting agents. However, these simpler subspaces represent a large loss of information; the consequence of this is that other populations strongly influence the policies obtained by using these state projections.
The result is that agents evolve to partner well with a broad range of other agents, rather than evolving to form optimal partnerships [ 46 ]. Thus, in addition to trying to decrease the complexity of the learning process, research in coevolution looks to achieve optimal policies rather than stable ones. Cooperative coevolutionary algorithms is a natural approach in domains where agents need to develop local solutions such as subsystem design , but the metric for success or failure is related to overall system performance [ 47 ].
In CCEAs, distinct populations evolve simultaneously, and agents from these populations collaborate to achieve good system solutions. One issue with CCEAs is that they tend to favor stable solutions, rather than optimal solutions [ 48 ]. This phenomenon occurs because the different evolving populations adapt to each other, rather than adapting to form an optimal policy. A further concern that appears with CCEAs is the problem of credit assignment. Since the agents succeed or fail as a team, the fitness of each agent becomes subjective and context-dependent.
In this case, the good agent may be perceived as bad [ 48 , 49 ]. This means that any action an agent takes which increases the value of the difference evaluation also increases the value of the overall system performance. This property is termed alignment. Also, note that the second term in Eq. This reduces noise in the feedback signal, meaning that difference evaluations are highly sensitive to the actions of an individual agent. In addition to the theoretical properties of alignment and sensitivity, difference evaluations have been proven to increase the probability of finding optimal solutions in cases where the optimal Nash equilibrium is deceptive [ 49 ].
Difference evaluations do not affect the location, number, or relative ordering of Nash equilibrium [ 49 , 53 ]. A deceptive Nash equilibrium is one in which the payoff drops significantly if one agent deviates. The actions associated with a deceptive Nash equilibrium correspond with low payoffs unless all agents simultaneously select the correct action.
Softcover reprint of hardcover 1st ed. A key insight of this approach is that complex system design, undertaken with respect to a variety of design objectives, is fundamentally similar to the multi-agent coordination problem, where component decisions and their interactions lead to global behavior. Your review has been submitted and will appear here shortly. Comparing the effects of weight variation between systems' objectives allow us to compare the design decisions of the agents. The system objective must also consider the cost of manufacturing and assembly operations.
In these cases, one agent deviating from the optimal strategy results in a large decrease in the overall system payoff, meaning that finding these Nash equilibria is typically extremely difficult. This paper demonstrates that, in a design problem, a team of autonomous agents can be used to design a complex engineered system. The first step is to define the design process for the agents as shown in Fig. To begin with, it is necessary to define the system-level objectives and the system constraints.
Then we select the team of agents, where each agent will be responsible for optimizing a specific subsystem. Second, using the different system-level objectives, it is necessary to define the overall system objective. The overall system objective will be used by the algorithm to measure the impact of the design concept for each agent team. Using cooperative coevolutionary algorithm, the agents will evaluate all the feasible combinations of solutions and choose the best one. In this paper, we will compare the final design of a system designed by a team of engineers against the design reached by a team of autonomous agents.
This paper will illustrate the proof of concept of the approach using the design of a Formula SAE racing vehicle. Formula SAE is a collegiate design competition that requires students to design, build, test, and to compete with a racing automobile [ 54 ]. Formula SAE works as a fictional company, where teams of students are contracted to create and build a functional small formula racing vehicle.
The final design is tested based on a series of rules, which ensure the safety of all operations and promote a design challenge for engineers. The objective is to create a racing vehicle, which will win the acceleration and skid-pad event of a Formula SAE race. The acceleration event evaluates the vehicle acceleration in a straight line on the flat pavement. The course layout for the acceleration event has a length of 75 m and 4. The skid-pad event measures the car's cornering ability on a flat surface while making a constant radius turn. The course layout for this event has two pairs of concentric circles in a figure of eight patterns.
The centers of these circles will be The inner circles will be In this paper, we compare the design process of a Formula SAE engineering team against a team of autonomous design agents. The design process for the autonomous agents will follow the same design principles as the one followed by the team of engineers, using the parameters in Fig. We will set some customer requirements for the vehicle performance. Second, we will define the system-level objectives and constraints.
Finally, the autonomous design teams agents will be defined, and an algorithm will be implemented. The key factor of the presented design process will be the design agent's selection. For the purpose of this paper, the system to be analyzed is going to be simplified. The selection of components such as the differential, clutch, and transmission is not being considered for this first part of the research. All the subsystems and components mentioned will be implemented as part of future work.
The first step is to investigate the form of a system-level design objective that ensures an intrinsically dependable and adaptable system directly from the design process. The key principle here is that the objective function should capture the designer's agent's underlying preferences for the system while ensuring the design is both adaptable and dependable.
The presented approach allows adaptability or learning within the design agents, the final design will have a significant performance improvement in comparison with other approaches that do not [ 55 ]. There are two key points related to adaptability. First, the system designer can change the overall system objective function by either adding new metrics or changing the weights of existing metrics, and the system can adapt its design to optimize the new metrics.
In this case, the system is adaptable to the performance desires of the system designer.
Second, the system can adapt to the design of different components being changed because the interaction between agents ensures that agents optimize their components not only with respect to local performance measures but also on how their components interact with the system as a whole. To win the competition, the system needs to maximize the vehicle acceleration and aerodynamic grip, and minimize weight, drag, and the location of the center of gravity. Engineering teams are responsible for designing the system, and each team is responsible for designing one specific subsystem as shown in Fig.
Consistent with the philosophy of engineering design, the goal is to make the decisions as accurately as possible without having to build prototypes or conduct costly testing. There are eleven design teams agents each responsible for a subsystem in the overall design problem.
The design agents in this system do not share any design variable. There are two key challenges with ensuring coordination, including cases where design agents do not share variables among themselves. First, the utility function for each subsystem being optimized must be constructed such that they ensure that the overall system is optimized.
For example, the rear wing agent rw may optimize the downforce of the system by making a larger wing.
Although the large wing would optimize the downforce of that specific subsystem, it may harm the overall system performance due to the increased weight and drag of the large wing. Thus, the agent should be optimizing the subsystem with respect to its impact on the overall system, not based on local performance measures. Second, when decomposing the design problem into multiple subsystem optimizations, each subsystem must be optimized in a manner that ensures it will work well with the additional subsystems being optimized. The agent optimizing the suspension must thus optimize the suspension such that it will perform well with the other subsystems being simultaneously optimized.
These teams, as well as the design parameters they are responsible for, are given in Table 1. Further, note that each team name has an abbreviation given in Table 1 to define variable naming conventions. So, for example, the height of the rear wing is denoted h rw. For the purpose of this analysis, the customer requirement for the designed vehicle is to win the acceleration event of a Formula SAE race.
The following assumptions will be used as the requirements for the system-level objectives and the environment:. We now discuss the objectives to be optimized for the entire system in Secs. Note that as there are two side wings, two rear tires, and two front tires, these mass values are doubled in the overall mass calculation. The third and fourth objectives are to minimize the overall drag of the vehicle and to maximize the downforce of the vehicle.
We assume that the components which influence drag are the rear wing, front wing, side wings, and cabin. We also assume that only the wings influence vehicle downforce. We will first analyze the wings and then the cabin. For the cabin, we assume a drag coefficient C d , c of 0. The eighth objective of the design process is to maximize the velocity of the car V cor while turning in the skip pad.
We assume that the forces which influence the radial acceleration of the car are the the overall downforce F d , the overall rolling resistance R roll , and the suspension force F sp. We assume that the front tires are in contact with the step simultaneously. In a Formula SAE competition, the car prototype is judged in a number of different events. In this paper, we are not replicating a Formula SAE competition; however, it is necessary to judge the design of the vehicle.
A weighted linear sum is our approximation on how to judge the design with respect to its performance. Intuitively, the difference evaluation compares system performance with and without agent i , to approximate the agent's impact on overall system performance. The constraints used for the vehicle were set according to the rules of the Formula SAE rules [ 54 ].
The SAE rules present the competition regulations technical and design requirements. The Formula SAE rules were used to define the minimum and maximum dimensions and the areas where the structural components are allowed to be placed. The CCEA is a stochastic, population-based search algorithm which is capable of exploring any design solution in the search space. Carliss Y. Harvard Business School.
Editors: Braha, Dan, Minai, Ali A., Bar-Yam, Yaneer (Eds.) Every time that we take money out of an ATM, surf the internet or simply turn on a light switch, we enjoy the benefits of complex engineered systems. And as the demands that we place on these systems become increasingly. Complex Engineered Systems: Science Meets Technology (Understanding Complex Systems) [Dan Braha, Ali A. Minai, Yaneer Bar-Yam] on cydyqywyty.cf
HBS Home. Business and Environment Business History Entrepreneurship.
Finance Globalization Health Care. Technology and Innovation. Finance General Management Marketing. Technology and Operations Management. Print Email. Citation: Baldwin, Carliss Y. About the Authors Carliss Y. Kim B. Baldwin The purpose of this chapter is to examine the value structure of flow production processes and to explain why it is necessary to rationalize flow processes using the tools of systematic management. I first explain the problems facing managers of multi-step flow production processes at the end of the 19th Century.
I introduce a model of the value structure of a production process made up of interdependent steps and define the production bottleneck of the process. I use the mathematical definition of a bottleneck to derive two general properties of stochastic multi-step flow processes with bottlenecks. These properties imply a need for ongoing managerial oversight and intervention using a set of tools that went by the label systematic management. Without active systematic management to address production bottlenecks, large-scale flow processes can easily collapse into chaos.