تئوری و آزمایش های طراحی سیستم های هوشمند تعاونی

In this paper, we identify the business problems that lend themselves to the design of cooperative intelligent systems and empirically demonstrate the design and application of a multi-agent intelligent system for production scheduling. Our experiments suggest that a multi-agent system where agents coordinate their actions generally performs better than a multi-agent system where agents do not coordinate their actions.

Cooperative intelligent systems is one of the dynamic research areas in information systems [3] and [29]. Several researchers have shown that cooperative intelligent systems can be used for low maintenance cost decision support applications that allow decision-makers to take complex decisions [29]. However, not all business problems are suitable for the design of cooperative intelligent systems and the design of cooperative intelligent systems requires consideration of several factors. Cooperative intelligent systems are also called distributed artificial intelligence (DAI) systems [25]. DAI consists of two sub-fields distributed problem-solving (DPS) and multi-agent systems (MAS) [5]. A DPS system consists of a set of independent geographically dispersed computer systems (problem-solvers) that share “solutions” to solve a problem that none of the independent computer systems can solve independently. Both data and knowledge, in a DPS system, are geographically dispersed. Unlike the distributed database systems and the inter-organizational systems (IOS), a DPS system shares solutions and not data. In a DPS system, problem-solvers that will work together on a given problem and the solutions that will be shared by problem-solvers are usually known. A multi-agent system (MAS) is a set of geographically dispersed computer systems that “dynamically” work together, through communication, to solve problems that none of the independent computer systems can solve independently. Unlike a DPS system, computer systems that will work together to solve a given problem are not known in a MAS and are decided dynamically. Further, communication among computer systems in a MAS can be data, hypotheses and knowledge. Multi-agent systems are of considerable complexity with respect to their functionality and structure [43]. For most application tasks, even in simple environments, it is difficult to determine the behavioral repertoire of an agent in a multi-agent system [43]. For example, determining behavioral repertoire of an agent in a multi-agent system requires a decision-maker to have a priori knowledge of future environmental requirements, knowledge of the availability of each agent at each environmental state in the future and knowledge about how agents will interact in response to the future environmental requirements [43]. The lack of availability of a priori knowledge necessitates the design of an adaptive system that can react to uncertain dynamic situations. A MAS offers features such as parallelism, robustness and scalability, which cannot be handled by centralized systems [43]. In particular, a MAS is used in domains that require integration of knowledge from multiple sources, resolution of different interests and goal conflicts [43]. Learning coordination in a MAS requires that agents adapt, adjust and learn to work with others agents to solve problems. The key issues of learning effective coordination in a MAS are the information exchange scheme and the coordination strategies between loosely coupled agents to achieve effective overall system performance. Given the complexity of designing a MAS, the current research aims to address the following issues. (1) What types of business situations are suitable for deploying a multi-agent system? And, (2) what are different design considerations related to a multi-agent system design? We answer both of these questions through literature review and an empirical demonstration of a multi-agent intelligent system for a production scheduling environment. The rest of the paper is organized as follows: in Section 2, we review available literature in cooperative intelligent systems (DPS and MAS) area. In Section 3, we describe a production scheduling problem, a framework for genetic algorithm (GA)-based learning and different types of coordination groups, and propose a general multi-agent coordination strategy. In Section 4, we detail the results of our simulation experiments and statistical analyses. In Section 5, we conclude the research by describing significance of our research and highlighting possible future extensions.

The focus of our research was to identify which types of business situations lend themselves to the design of cooperative intelligent systems and empirically demonstrate the design and application of multi-agent system's application for manufacturing. Using literature review, we developed a set of criteria that may be used to identify business situations that are suitable for the design cooperative intelligent systems. We then embarked on the design of a multi-agent system for manufacturing flow shop scheduling. We proposed a coordination scheme for a multi-agent system and tested it against an existing framework of multi-agent learning without coordination. The coordination scheme operated through a coordinating agent. The coordinating agent coordinated its actions with the other dependent agents by exchanging taxes. A tax was a percent of the total payoff that a dependent agent gave to the coordinating agent for coordinating its activities with the dependent agent. The total payoff was the combined performance of the agents for a given activity. The performance was determined by the managerial objective of the multi-agent framework. Three basic coordination groups were proposed and used in our study. The three different types of coordination groups were inward fork, outward fork and straight-line. The three different coordination groups corresponded to three different types of dependencies between the agents. In an inward fork coordination group, many to one relationship was modeled. There were two agents that coordinated their actions with one coordinating agent. In an outward fork coordination group, one to many relationship was modeled. There was one coordinating agent that coordinated its actions with two dependent agents. The straight-line coordination modeled one to one relationship with one coordinating agent that coordinated its actions with one dependent agent. A generalized framework of coordination was applied to a manufacturing shop floor environment. The agents in the manufacturing shop floor were the dispatchers that dispatched jobs over a machine. The agents contained a knowledge base of dispatching rules and a genetic algorithm that learns new dispatching rules over time. Genetic learning worked through the survival of the fittest principle. The fitness of the rules, upon their evaluation, was determined by the performance objectives. The performance objectives that were considered here were minimizing flow time of jobs and minimizing tardiness of jobs. Separate experiments were conducted for each of the individual payoff functions. Two different sets of experiments were conducted. The first set of experiments involved testing the coordination strategy for three different types of coordination groups under ideal conditions with no breakdowns in machines. The second set of experiments involved testing coordination for a realistic larger configuration [4] with machine breakdowns. The results of the experiments indicated the overall utility of coordination. At 0.01 level of statistical significance, differences were found between the performance of no-coordination-based learning and coordination-based learning in all cases except for outward fork coordination group with minimize flow time objective and no machine breakdowns. The inconsistency in results with regard to outward fork for flow time objective may be explained by the basic difference between the structural representation of outward fork coordination group and the inward fork and straight line coordination groups. The results of our study are generalizable for the conditions set in our experiments. However, much can be done to improve the generalizability and results of our study. For example, in the current research, a “prescriptive”, rather than a “descriptive”, approach to multi-agent coordination was used. The agents coordinated to maximize their own individual utilities. Some DAI researchers were critical of using the individualistic approach to designing systems. Gasser [15] said “society comes first” and opposed the individualistic and psychological approach to DAI systems. AI researchers have long considered that AI was concerned with individual mind (individual intention, action, etc.) and DAI was concerned with social actions (collective mind, intention, etc.) and mental states which are shared, joint, and sometimes collective [27]. From the critical standpoint, the current research ignored social relations and long-term commitments. Issues such as coordination and dependence that were independent of agent awareness and choice were not considered in the current research. These issues need consideration in the future if the “social” view of the agent was to be adopted. Given that the collective group performance relied upon the objective structure of interdependence among individual agents, our research approach (individual intention, action, etc. of agents) offered several advantages over adopting the “social” view of the agent. Some of the advantages were: 1. Agents do not imply mutual belief of the overall plan. 2. Mutual dependence between dependent agent and cooperative agent (as both share the common goal) ruled out the situations of competition [10] and [18]. The common goal assumption helped resolve the conflict (tax structure) between the agents as the social norm appears to be “higher taxes should result in higher future overall payoff” and “no to very limited cheating between the agents (breakdowns)”. This norm may cease to exist when agents do not share common goals and/or have their own local goals. Local goals may induce competition and conflict between the agents in the current framework and conflict resolution using the current coordination scheme might be limited. Future work needs to be done in investigation coordination where agents in multi-agent framework do not share common goal. Most of the experiments in the current research assumed that the belief about mutual dependence holds true. Under the assumption of mutual dependence, each agent believed that it depends on other agents to achieve its goal. The mutual dependence fostered cooperation between the agents. In most cases, under the assumption of mutual dependence, coordination worked rather well and no misunderstandings occurred. In the case where some of the agents failed to perform (machine breakdowns), the mutual dependence assumption was violated. A dependent agent was cheated (as a result of coordinating agent breakdown). The question of cheating raises a number of questions such as, can a coordination group be efficient if some of the agents spend all their resources without being adequately rewarded or without being able to achieve their own goals? How can a self-interested agent decide to do something for the group, which may cost it more than it gains? How does an agent cope with cheating (machine breakdowns)? Are social norms that prevent cheating required? (this may need to change an individualistic agent to approach to a “social agent” approach proposed by critics). It appears that the mutual dependence assumption was the single most important success factor of the proposed coordination scheme. The mutual dependence assumption with limited or no cheating encouraged the agents to coordinate and resolve conflicts (in terms of adjustment of taxes) and speed up the convergence. Serious violations to the mutual dependence assumptions might be a potential pitfall for the current coordination scheme and need future investigation. Some of the avenues for the extension of the current coordination scheme are: (1) redesign the coordination scheme that penalizes an coordinating agent for cheating (breakdowns) and (2) develop a set of norms regulating the coordination between the agents in a way that is beneficial for the whole group and not for any agent in particular.