Chapter 4 : Applications of Distributed Artificial Intelligence in Industry

In many industrial applications, large centralized software systems are not as effective as distributed networks of relatively simpler computerized agents. For example, to compete effectively in today's markets, manufacturers must be able to design, implement, reconfigure, resize, and maintain manufacturing facilities rapidly and inexpensively. Because modern manufacturing depends heavily on computer systems, these same requirements apply to manufacturing control software, and are more easily satisfied by small modules than by large monolithic systems. This paper reviews industrial needs for Distributed Artificial Intelligence (DAI),1 giving special attention to systems for manufacturing scheduling and control. It describes a taxonomy of such systems, gives case studies of several advanced research applications and actual industrial installations, and identifies steps that need to be taken to deploy these technologies more broadly. 1. The Demand for Multi-Agent Systems This section describes the vision of agile manufacturing, outlines challenges that conventional CIM systems face in reaching toward this vision, and suggests why systems of autonomous agents may be particularly well suited to overcoming these challenges. 1.1. The Challenge of Agility2 Agility, the ability to thrive in an environment of continuous and unanticipated change, provides competitiveness in global markets. Manufacturers with shorter product cycles can track customer desires more closely and thus capture market share. Firms that can change volume of production rapidly can avoid sinking large amounts of capital in excess inventory, while still maximizing returns in periods of rapidly increased demand. With the demise of the cold war, agility is also the foundation for dual-use strategies that permit high levels of preparedness without diverting industrial investment into dedicated defense facilities. Table 4.1 compares agility with other manufacturing objectives along four dimensions: cost, time, quality, and scope. 1In the sense used here, DAI systems may include simple reactive modules that would not individually be considered artificially intelligent, as well as humans integrated electronically into the overall system. 2This section relies heavily on (Dove 92). Cost Response Time Quality Scope Relation to Agility Agile Low Quick Self Healing Limitless Flexible Low Fast Strong Limited Superseded Lean High Slow Fragile Little Complemented Table 4.1: Characteristics of Agility Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 2 Agility requires systems that are low in cost, quick to respond, able to correct themselves in the presence of faults, and unlimited in scope. The first two characteristics have been explored in the context of flexible systems (systems engineered to switch rapidly among predefined members of a family of products), but only within predefined limits, and quality has been guaranteed by brute strength rather than the ability for self-correction. Lean manufacturing invests heavily, both in time and money, to obtain a facility that is optimized for a particular task but deteriorates rapidly outside the design envelope. Agility impacts the entire manufacturing organization, including product design, customer relations, and logistics, as well as production. For the sake of concreteness, this discussion focuses on production systems, but occasionally mentions applications of multi-agent systems to other manufacturing functions as well. 1.2. Problems with Conventional CIM Systems. Conventional systems for Computer-Integrated Manufacturing (CIM) are often highly centralized. At the least, a central DB provides a consistent global view of the state of the enterprise, both for internal reference and as an interface to the rest of the organization. Typically, a central machine also computes schedules for the facility, dispatches actions throughout the factory in keeping with the schedule, monitors any deviations from plan, and dispatches corrective action. This approach has three characteristics that impede agility. It is centralized; it relies on global plans; and it precedes execution with planning and scheduling. These characteristics aggravate four operational challenges to manufacturing scheduling and control (MSC): stochasticity (the impact of noise and randomness, such as machine breakdown), tractability (the combinatoric impact of large numbers of interacting entities), decidability (the operational implications of the factory's formal equivalence to a Turing machine, making most of its interesting properties undecidable), and formal chaos (sensitivity to initial conditions resulting from nonlinearities) (Parunak 91), as outlined in Table 4.2. A centralized MSC approach is especially susceptible to problems of tractability, because the number of interacting entities that must be managed together is large and leads to combinatorial explosion.3 All four challenges to MSC make it difficult to predict what will happen, and thus to form a global detailed plan that can be executed after it is computed. The need for sequential planning and execution is particularly difficult in the face of stochasticity (which changes the state of the world in ways unanticipated in the plan and thus makes the plan invalid) and tractability (which increases the computational resources needed to derive a schedule, to the extent that there may not be enough time available to complete a schedule before it is needed). In addition to these operational challenges, it is difficult to implement agile systems with conventional technology in the first place. Centralized software is large, complex, and expensive. It is difficult to bring on-line. The resulting credo, "If it ain't broke, don't fix it!", is a strong disincentive to change. Because it requires extensive customization for each installation, it is difficult to realize economies of scale that would make new installations less expensive and easier to bring on-line. 1.3. Conventional MSC Approaches The major technologies used today for MSC, particularly for scheduling and planning, are heuristic dispatching verified by simulation, numerical programming, traditional AI, and advanced heuristics. 3Some multi-agent systems rely on some sort of central control to coordinate their agents, often through a NASREM-type hierarchy (Albus et al. 87). While this approach lends itself well to conventional control techniques, the hierarchy induces a strong binding among agents that makes agile reconfiguration difficult. Centralized Planned Sequential Stochasticity X X Tractability X X X Decidability X Formal Chaos X Table 4.2: Challenges to Traditional MSC Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 3 1.3.1. Heuristic Dispatching For many practical applications, shop floor control is dominated by heuristic dispatching, in which a simple decision rule determines the next job to be processed at a given workstation. Typical rules are "select the job with the shortest remaining processing time," "select the job with the earliest due date," and "process jobs in the order received.". Dozens of simulation studies have examined the performance of hundreds of such rules with respect to parameters such as total cost per job, number of late jobs, machine and labor utilization, and mean and variance of job residency in the shop (Pawalkar & Iskandar 77). The simplicity and robustness of dispatching rules makes them attractive in the real world. However, there are regions of the solution space of schedules that cannot be generated by applying a single dispatching rule to every workstation in the plant (Blackstone et al. 82). 1.3.2. Numerical Programming Techniques such as linear and integer programming offer clean analytical formulations of scheduling problems, and the combination of new algorithms and more powerful hardware make them feasible for solving fairly substantial problems. Their great advantage is that they can find true optima. The disadvantage is that they are combinatorially explosive and computationally intensive, and so require either a large investment in supercomputers or an environment that can wait for slow answers. Thus their main value is in relatively static situations such as long production runs. Even there, plans can be rendered useless by equipment breakdowns or the arrival of a priority order. Because of the time required to schedule a large facility completely by numerical programming methods, two techniques have been developed to permit partial rescheduling: DEDS (Ho & Cao 83, Ho & Cassandras 83) and turnpike theory (Bean & Birge 85). DEDS (Discrete Event Dynamic Systems) adopts the techniques of perturbation analysis from continuous systems to the requirements of a discrete environment. Once a nominal trajectory of the system has been obtained (by simulation or exhaustive scheduling), limited deviations from the trajectory can be analyzed without complete recomputation. Thus predictions can be corrected within certain limits without recomputation. Turnpike theory derives its name from the problem of a traveler who has detoured off the highway before nearing the destination. Under certain conditions, it is easier and cheaper to find a way back to the highway than it is to replan the entire trip from the current location to the destination. Applied to scheduling, the challenge is to find ways to return to a previously computed schedule. DEDS and turnpiking can help reduce rescheduling for a large facility that drifts off schedule, but are of relatively little use in the kinds of reconfiguration and capacity shifts anticipated in agile manufacturing. Agility requires machine configurations that can be changed daily and batch sizes approaching unity, not just small incremental shifts from present practice. The need to tailor the program to the facility being scheduled, as well as the time-consuming scheduling process, both make numerical programming less than ideal for an agile environment. 1.3.3. Traditional AI Traditional AI has devoted considerable attention to problems of manufacturing scheduling and control (Smith 91). By taking into account semantic information about the domain that does not lend itself to numerical computation; by applying heuristics judiciously and selectively (rather than globally as with dispatch rules), and by adopting a "satisficing" approach that does not insist on a theoretically perfect optimum, symbolic computation has led to systems that are somewhat faster than numerical programming and are more flexible and able to accommodate richer constraints, while yielding results superior to dispatch rules. However, these systems still tend to be large, complex, and specific to a particular installation, thus making them expensive to construct and difficult to maintain and reconfigure. Furthermore, while they are faster than some numerical programming codes, they are not fast enough for a facility whose configuration and load changes daily. 1.3.4. Advanced Heuristics Recent research in operations research (Morton & Pentico 93) combines heuristics, simulation techniques, and mathematical optimization theory in various ways to address scheduling problems. Such techniques overcome the restrictions of simple heuristics, and have many characteristics in common with AI approaches. Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 4 Implementations with an industrial track record are successful in some but not all contexts (for example, OPT (Meleton 1986, Lundrigan 1986, Vollman 1986) has difficulty with shifting bottlenecks), and some of the newer theories are still untested in industrial conditions. 1.4. How Can Multi-Agent Systems Help? This section briefly reviews the characteristics of multi-agent systems and, based on the business context outlined above, suggests the kinds of applications where they are most likely to be valuable. 1.4.1. Characteristics of Multi-Agent Systems Multi-agent systems offer a way to relax the constraints of centralized, planned, sequential control, though not every multi-agent system takes full advantage of this potential. They offer production systems that are decentralized rather than centralized, emergent rather than planned, and concurrent rather than sequential. The autonomous agent approach replaces a centralized database and control computer with a network of agents, each endowed with a local view of its environment and the ability and authority to respond locally to that environment. The overall system performance is not globally planned, but emerges through the dynamic interaction of the agents in real-time. Thus the system does not alternate between cycles of scheduling and execution. Rather, the schedule emerges from the concurrent independent decisions of the local agents. Autonomous agent systems are inspired by models from biology (ecosystems) and economics (markets), in contrast with the military patterns of hierarchical organization manifested by traditional approaches. Table 4.3 contrasts some of the advantages and disadvantages of the two philosophies. On the one hand, the autonomous agent approach may face some disadvantages. Theoretical optima cannot be guaranteed. Predictions for autonomous systems can usually be made only at the aggregate level. In principle, systems of autonomous agents can become computationally unstable. The degree of seriousness of these challenges needs to be assessed empirically: the optima computed by conventional systems may not be realizable in practice, and the more detailed predictions that conventional approaches permit are often invalidated by the real world. On the other hand, an autonomous approach appears to offer some significant advantages over conventional systems. Because each agent is close to the point of contact with the real world, the system's computational state tracks the state of the world very closely, without need for a centralized database. Because overall system behavior emerges from local decisions, the system readjusts itself automatically to environmental noise or the removal or addition of agents. The software for each agent is much shorter and simpler than would be required for a centralized approach, and as a result is easier to write, debug, and maintain. Because the system schedules itself as it runs, there is no separate scheduling phase of operation, and thus no need to wait for the scheduler to complete. Issue Autonomous Agents Conventional Model Economics, Biology Military Issues favoring conventional systems Theoretical optima? No Yes Level of prediction Aggregate Individual Computational Stability Low High Issues favoring autonomous agents Match to reality High Low Requires central data? No Yes Response to change Robust Fragile System reconfigurability Easy Hard Nature of software Short, simple Lengthy, complex Time required to schedule Real time Slow Table 4.3: Agent-Based vs. Conventional Technologies Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 5 1.4.2. When should Multi-Agent Systems be used? Agent-based systems offer the greatest promise in applications with several characteristics. Distribution.--Agents are an inherently distributed mechanism, and are a promising strategy when other constraints make a centralized architecture undesirable. Thus they are the technology of choice for interaction among widely distributed shops, or when for other reasons a centralized control computer is undesirable in a single shop (for example, when the single point of failure it presents is not acceptable). Factorization.--Much of the power of an agent comes from its identification with some entity (such as an information source, a machine, a part, or a tool) that makes sense in the application domain. When the problem is easily conceived in terms of such naturally-occurring entities, agents can be applied fairly easily. However, factorizations that are suggested by traditional analysis but do not correspond to naturally occurring entities (such as the hierarchical decomposition of a factory) can lead to very inefficient agent architectures. Variability.--A system whose parameters do not vary widely can be optimized through a carefully planned traditional scheduling and control system. A population of agents can reconfigure itself as it runs, an important advantage for systems that must respond to a wide range of different conditions. For example, a classical approach may be competitive in a paint shop for a factory that makes only one color of car, but a case to be described below shows the benefits of an agent architecture when many colors must be supported. Change.--Using conventional techniques, the most expensive part of a manufacturing system is not the machinery, tooling, or energy to operate it, but software creation and maintenance. When a system has a long lifetime, this expense can be amortized over many years. When systems are expected to change frequently, the agent approach shifts the software effort from integrated systems that depend on a particular configuration to the individual agents that can be swapped in and out as the entities they represent are shuffled around. 2. A Taxonomy of Multi-Agent Systems Several taxonomies of multi-agent systems have been published (Bond and Gasser 1988, Decker et al. 1989, Demazeau & Mueller 1990, Durfee et al. 1989, Grant 1992, Huhns 1988). The taxonomy presented here has been developed from these, giving special attention to features that are most relevant from an application perspective, based on an extensive survey of applied research and development in industrial applications of MAS. (This paper does not cite all of the cases considered, but only representative examples.) Three groups of perspectives are important: the manufacturing function to which MAS technology is applied, the architecture of an individual agent, and the architecture of the system within which agents interact. 2.1. Application Function In manufacturing, most MASs are applied to two functions: production and design. By far the most common application of MASs to manufacturing is in production, mostly in scheduling and control and to a lesser extent in monitoring and diagnosis. The fit of such systems to production lies in their ability to represent the distributed entities on the shop floor and endow each with some level of local intelligence. Several systems (e.g., (Klein 1991)) support the design function. Here, multiple agents help designers to interact with one another and with their production counterparts to avoid design conflicts and ensure manufacturability. A few systems (e.g., ARCHON/GRATE (Wittig 1992, Cockburn & Jennings 1994)) deal with other manufacturing functions, such as power distribution , information integration, and logistics. 2.2. Individual Agent Architecture In considering individual agents, we consider how similar they must be with one another into order to interact, and how sophisticated the reasoning is within each agent. Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 6 2.2.1. How diverse are agents from each other? If agents are to interact with one another, they must have something in common. In most research systems, agents have differing functions, but share a common architecture. Sometimes this commonality extends to every aspect of an agent except for a few parameters that it manipulates (Morley & Schelberg 1993). At the other extreme are "body-head" architectures, where agents' bodies can have radically different architectures, even being different off-the-shelf products, but with a common "head" to permit communications among agents (Jennings et al. 1992). (In the simplest case the "head" takes the form of a standard language, such as SQL.) "Body-head" architectures are a common mechanism for incorporating humans as peer agents in a system. Systems whose agents differ from one another by more than parametric variation may be termed "heterogeneous." While communications usually takes the form of digital messages over a network, this is not the only possible coordination mechanisms. Agents may coordinate solely through their effects on a shared environment. In this case, the "head" of a "body-head" architecture consists of common sensor-effector modalities. 2.2.2. How sophisticated is an individual agent's reasoning? Different levels of sophistication can be used by different agents, or even by the same agent at different times and under different circumstances. (Demazeau & Mueller 1991) offers a variety of studies along this important dimension. Levels of Sophistication In theory, an agent's sophistication can range from a simple sensing agent that reacts to its environment but has no memory and no model of other agents, all the way up to full human capabilities. In practice, actual implementations add at least memory to sensing, so that the agent maintains local state. The next level of sophistication is self-consciousness, in which each agent knows of the existence of other agents distinct from itself, and thus can carry on rudimentary communication. A social agent goes a step further and models other agents' states, goals, and plans. Higher capabilities yet include such functions as as making commitments to one another, planning tasks, and learning from experience (Shoham 1993). Few industrial applications of artificial agents at present go beyond self-conscious agents. In industry, more complex functions are usually provided by using artificial agents as an interface for a human operator, to whom they furnish information and from whom they take commands. Thus the main point of the system is to augment, rather than replace, the human operator. From an industrial perspective, it is both expensive and technically complex to automate the controls on much existing production equipment to the point that a human operator could be replaced, and social and political concerns also challenge the wisdom of eliminating the human. Thus the most direct route for many firms to using DAI on the factory floor may be via products that are emerging under the rubric of "manufacturing execution systems" (AMR 1993, Gillespie 1992). These systems provide human staff throughout a manufacturing enterprise with electronic connectivity, data access, and decision support, and may be considered "groupware for the shop floor." Combining Reaction and Planning In some systems, the various levels of sophistication within each agent are separately accessible. For example, sensing capabilities provide a rapid reflex response, while higher level planning capabilities are invoked if there is time. The term "subsumption architecture" for this technique was introduced by (Brooks 1986), who eschews any symbolic representation within an agent (Brooks 1991), but the term can be extended to apply to architectures that include symbolic representations as well. When we do add more sophisticated reasoning capabilities to a reactive agent, we would like to do so in a way that does not forfeit the benefits of simple reaction. Systems for reactive planning fall along a continuum from pure reaction (which is fast but inflexible) to pure planning (which is slow but highly adaptive) (Table 4.4). Parunak: Applications of Distributed Artificial Intelligence in Industry 04/08/94 10:04 AM Copyright  1994, Industrial Technology Institute. All Rights Reserved. Page 7 The top three cases fall in the "reactive planning" category. In pure reaction (Figure 4.1), there is no way to alter the behavior of the reactor. No planning takes place at run-time. The system is inflexible, but fast. (Brooks 1986, 1991) Such an architecture is most appropriate in tasks where the overall envelope of possible tasks is understood well in advance. Brooks applies the model to a mobile robot, and it would be appropriate for intelligent mobile vehicles used to transport materials around a factory. In reaction overridden by plan (Figure 4.2), the planner monitors the external world in parallel with the reactor, and when it disagrees with the reactor, pushes its own instructions out to the actuators, thus overriding the reactor for the nonce. However, the next time the situation arises, the reactor will still react the same way. (The planner may cache the preferred response so that the reactor can be overridden more efficiently the next time.) Because the preferred response comes via the planner rather than the reactor, it is slower. (Cohen et al. 1989) The developers of this model apply it to unmanned bulldozers used in fighting forest fires, emphasizing its suitability where the domain presents considerable uncertainty. It would be useful for agents that support interaction of human designers. Reactions can take care of anticipated requirements, while the system can plan around unanticipated situations without jeopardizing system security. In reaction modified by plan (Figure 4.3), the planner can rewire the reactor in real time so that in the future a new behavior will be performed at reaction speeds. This rewiring is in software, thus reactors of this class may be slower than hardwarewired reactors a la Brooks, but the response is still faster than invoking a planner. (Lyons & Hendriks 1992) This approach is useful for agents that must handle uncertainty in real-time, such as applying emergency corrections to system failures. 2.3. System Architecture Individual agents require the support of an overall system within which to interact. This system must provide a means for agents to communicate with one another, and define protocols for the resulting conversations. The special case when some agents are artificial and others are human merits special consideration. 2.3.1. How do agents communicate with each other? The notion of a multi-agent system presumes that individual agents can change their shared environment, sense changes in this environment, and alter their behavior in response to these changes. These interactions via environmental change fall into two categories: performance and convention. Performance interactions are those defined by the environmental changes that an agent must make in performing its domain function, whether or not there are other agents in the system. For example, a manufacturing cell can only run when its input buffers contain parts and there is room in its output buffer. As observers, we may interpret the action of placing a part in an input buffer or removing one from an output buffer as a "message" to the Fastest Pure reaction Flavors (Morley & Schelberg 1993, Steels 1990, Brooks 1986) Reaction modified by plan RS (Lyons & Hendricks 1992) Reaction overridden by plan Phoenix (Cohen et al. 1989) Pure planning BOSS (Smith & Hynynen 1987) Slowest Table 4.4: From Reaction to Planning