Intelligent Multi-Agent Based Information Management Methods to Direct Complex Industrial Systems ()
1. Introduction
The current industrial systems are destined to become environments even more technologically advanced in every aspect. This is due to several reasons: the complexity of the new manufacturing processes, the need to optimize costs, the increasing of the complexity in managing resources, the need to improve performances, the aspects tied to the information exchanges, the manufacture of complex products, the need to ensure advanced safety standard, and so on. All these aspects imply a continuous and growing renovation of the current industrial activities. Artificial Intelligence (AI) techniques [1,2], utilized to overcome these new qualitative and quantitative challenges bring from the next industrial age, seem to provide the more suitable support since they naturally have all these technical features (e.g., adaptability, reasoning, learning) required from the new manufacturing processes [3]. In particular, the use of AI based strategies allow to the industrial environments to adopt autonomous, dynamic and intelligent procedures to face expected and unexpected issues. The IAs and MASs are technologies that can play an important role in every aspect involved in the development of the next generation of the advanced industrial systems. In fact, these systems have to be designed to include intelligent, autonomous and evolvable entities in turn composed by different sub-entities having the same features. The number of levels (i.e., entities, sub-entities) and the complexity of each entity and sub-entity depend on the specific kind of the industrial system. The mentioned features are strongly tied to the IAs and MASs paradigms [4,5]. Moreover, several IAs and MASs have been just adopted in a wide range of applications that are connected to advanced industrial systems, such as: Robotic, artificial vision, production planning, systems control, engineering, information exchanging, and so on. In what follows, we describe the main features of the IAs and MASs and how their technology can be adapted to support the current and next generation advanced industrial systems.
The paper is organized as follows. Section 2 introduces the general aspects of the logical architecture of the current industrial systems. Section 3 describes the main features of the IAs and MASs to be adapted to support industrial systems. Section 4 provides a study of how a MAS is utilized within a productive process. In particular, the section discusses the different MASs algorithms highlighting their use to support the real-time production process. Section 5 shows the principles of the hybrid systems and presents the main characteristics of a case study. Section 6 concludes the paper.
2. Current Industrial Systems Architecture
The logical architectures of the current industrial systems can be seen as a correlated set of connected crucial processes that have to achieve a prefixed objective taking into account the current elaboration state of every other process. All the processes are aimed to reach a common target: The Artifact. In our context, the last mentioned term has to be considered both a physical object (e.g., technical or mechanical device, industrial product) and an abstract object (e.g., general service, immaterial product). The concepts regarding the creation or delivery of an artifact can be generalized to provide a common description of current industrial system architectures.
Each process deals with a specific activity (or part of it) of the whole industrial process. In fact, a generic Industrial system includes heterogeneous activities, such as: economical planning, quality and test control, real-time production control, internal activities, processes monitoring, robot control, sales planning, and so on. For this reason, the specification of a process can be completely different from another. As shown in Figure 1, a process is an autonomous entity characterized by internal rules and protocols. Every process works in coordination with others by taking into account their internal state.
The internal rules and protocols are highly dependent from the specific process. In general, the output of a process (i.e., artifact and information) regards a collection of homogeneous activities, while the input (i.e., processes interaction) concerns the connection of several heterogeneous activities, such as: Technical and schedulling information, control and manufacturing activities, data, artifact (or part of it). Every single process is composed of several different tasks (or sub-tasks) concerning a specific part of the whole industrial process. As shown in Figure 2, a task (or a sub-task) is an activity that can be solved, by decisional processes, using autonomous and intelligent methods without a specific knowledge of the surrounding processes. The depth of the tree depends
Figure 1. A general scheme of a process.
Figure 2. Process, tasks and sub-tasks hierarchy.
on both the kind of process and its complexity. Usually, processes concerning long term economic planning and technical activities are more complex than others.
The process at the root level (i.e., level 0) provides each of its tasks the needed input to accomplish the related assignments. Likewise, each further task (i.e., level 1) provides each of its sub-tasks the suitable part of the original input to accomplish the related sub-assignments. This activity of splitting input and assignments ends at the last level of the tree. The solution of each task and sub-task, by a decisional process, can require information or activities coming from different tasks and/or sub-tasks connected to any tree level. For this reason, the interaction mechanisms are a critical aspect of the current industrial systems. A failure in a task can produce a critical crash of the related process, or of the connected processes. Furthermore, the last mentioned aspect is fundamental to establish the intelligent mechanisms to solve the heterogeneous activities belonging to the complex industrial systems (e.g., real-time monitoring, dynamic business planning, chain control, simulation environments) and to overcome the expected and unexpected issues (e.g., breakdowns, environment changes, damages).
The decisional process that solves a task at the level 1 of the hierarchy can interact with any other decisional process that solves any other task at the same hierarchical level. Subsequently, the decisional processes that solve sub-tasks on the others levels of the hierarchy (e.g., level 2) have to directly interact only with the decisional processes that solve a sub-task connected at the same task or sub-task at the previous hierarchical level.
In others worlds, on each tree level, only the “brothers” can directly interact. If decisional processes belonging to different “fathers” have to interact, then the same “fathers” will accomplish the interaction activity.
As shown in Figure 3, the artifact is the result of joined processes. Likewise the previously described cooperation strategy, each decisional process that solves a process within the level 0 can interact with any other decisional process that solves any other process at the same level 0. Also in this case, if two decisional processes, driving two different tasks (e.g., at level 1) belonging to two different processes, have to interact, the same two processes will accomplish the interaction activity. This kind of hierarchical interaction serves to avoid disorder (i.e., anarchy) between tasks, sub-tasks and processes. Moreover, it is useful to allow the decisional process to have a total autonomy with respect the fixed assignment.
Each decisional process involved in the solution of a process, task or sub-task has to have well defined features. These features specify the way (e.g., protocols of communication and execution, priority of the decisional tree, priority of the queues, relationships between tasks or sub-tasks) and the resources (e.g., technical, economical, time planning, supply chain management) to carry on the activities. A generic decisional process can be considered as a set of activities that can be carried on by an intelligent, autonomous and flexible entity. Technically, these entities can be seen as a set of interactive procedures able to learn and evolve according to their life cycle. For these reasons, the use of IA and MAS technologies are considered the best solutions to these kinds of issues. As shown in Figure 4, each decisional process should be driven by a single IA. Each IA knows only the rules, the procedures and the protocols related to its specific task. It has a limited point of view regarding the whole industrial process and it is able to adopt solutions only on its specific contextual environment. Moreover, the networked set of the IAs (i.e., MASs) should drive each process into the industrial system. This association between tasks, assignments and intelligent entities allows to implement independent customized solutions to solve global and local specific issues.
3. IAs and MASs in Industrial Systems
The agent and multi-agent technologies have been widely adopted in robotic and automatic control system fields. The growing complexity of the current industrial environments and their affinity to the mentioned fields has encouraged the design of IAs and MASs aimed to direct
Figure 4. Decisional process and intelligent agent.
each aspect of the industrial chain. The CI based techniques promise to be the best way to perform these historical changes. There are several similar definitions of intelligent agent [6,7]. Considering the set of these definitions, in our context an intelligent agent can be described as: A social, autonomous and intelligent computational entity capable to accomplish a definite objective through an adaptive reasoning. Regardless of how an IA is defined, it is characterized by several general features [8] that can be adapted to our context. These features, joined with the given definition, complete the IA paradigm.
3.1. IA Properties Adapted to Industrial Systems
There are five main properties belonging to the IA that have to be adapted to build a basic industrial oriented IA.
3.1.1. Reactivity
An IA is a reactive entity that continuously perceives and affects with its surrounding environment. In an industrial system this feature depends on the process on which the agent is applied. For example, in economical planning processes the reactivity can be considered a not active feature, i.e., the agent has a starting receptive state and it will tend to switch its condition only if it recognizes particular changes in the environment. Instead, in technical processes the reactivity can be considered active, i.e., each change in the environment is considered an important event that has to be analyzed in real-time.
3.1.2. Pro-Activity
An IA has to operate in an expected way, but it can also take initiatives to overcome unexpected events. Usually, this feature does not depend on the process, it is tied to the environment in which the agent has to operate. The environments into the industrial processes can be classified in deterministic and non-deterministic. At the first class often belong processes derived from economical tasks that are tied to several unexpected changes (e.g., economical plans, supply chain management). At the last class almost always belong processes derived from technical tasks that are usually scheduled and coordinated (e.g., test and control, robot interaction, real-time monitoring).
3.1.3. Autonomy
An IA has to be capable of autonomous actions to accomplish its objectives. There are several situations in which an agent has to autonomously decide what is necessary to do, and how it has to be done. The main issue of this feature is to decide which, how many and when specified actions/activities can be adopted by the IA. As general rule, an IA is authorized to take autonomous decisions only if there is the real possibility that the whole process crashes. These situations occur quite often in advanced industrial systems where several technological actors are involved. In fact, the introduction of complex systems in the industrial environments (e.g., real-time systems, on-line robot vision monitoring, automatic control management) has led high level of entropy. Usually, a reasoning decision tree is used to decide the actions that have to be performed [9,10]. By the autonomy property each agent can interact with the environments or other agents without an external presence to ensure the achievement of the prefixed objectives.
3.1.4. Flexibility
An IA has to interact with the surrounding environment in different ways. It has to have the capability to adopt quickly itself to drastic changes into the relationships, environments or events. Sometimes conventional communication or interaction ways could be not sufficient to reach a prefixed objective. For example, an agent could want data from another agent in different way respect the standard protocols. For this reason, an agent has to manage its characteristics (e.g., output, external interface, internal protocols). This property is particularly important in industrial systems because they are prone to dynamic changes in management and technical levels.
3.1.5. Social Ability
An IA has to interact with other agents (also humans). This feature is the core of the intelligent agent theory; through the interaction an agent can understand the events and adapt its characteristics to the dynamic situations surrounding it. An IA needs to communicate with the internal and external environments to ensure several main activities (e.g., find out the state of other agents, sub-tasks synchronize, tasks scheduling, acknowledgement). Furthermore, it has to be considered that some agents (e.g., mediator agents, contractor agents, negotiator agents) have particular assignments that, more than others, need to interact with the surrounding entities. In industrial systems this feature takes a key value in business plans and technical activities.
The last property introduces a main matter in the industrial systems, i.e., the languages through which the agents exploit their own social ability. In general, the most important challenges in Agent Oriented Software Engineering (AOSE) is to use or create a suitable Agent Oriented Programming Language (AOPL) to allow the agents an effective and efficient interaction. There are several advanced tools that aid the developers in designing complex agent based software systems. Moreover, there are several powerful programming languages designed to facilitate the programming of agent entities. The chosen of the correct tool and language to achieve a prefixed assignment is a critical matter. Independently from the agent implementation there are some issues (e.g., debugging, validation, verification) that depend on the specific context in which the agents have to operate.
In AOPLs the focus is on how to describe the behavior of an agent in terms of constructs [11] (e.g., plans, communication strategies, interaction patterns, goals, messages). This agent description is aimed to specify the “reasoning state” of the agent during its activities. A typical example is AGENT-0 [12] a simple and powerful generic agent interpreter. This AOPL tends to program agents in terms of basic behavior, in this way the properties of the agents can be developed by definite notions, such as: Reactions, communication ways, interactions, rules, and so on. Another interesting AOPL is dMARS (Distributed Multi-Agent Reasoning System) [13], which provides both a sophisticated monitor and manageable macros to develop interaction activity between agents. Moreover, this AOPL provides several functionalities to manage critical implementation steps (e.g., configurations of internal states, perception of events). The dMARS can be considered a reference technology to direct complex industrial systems since it has a wide range of high solutions that support the whole industrial processes. A last interesting AOPL considered to develop agents in industrial systems is JACK [14]. It is a versatile framework designed to create, by models, the concepts that drive the intelligent agents. This language has a wide use in industrial contexts since it provides powerful tools for the run-time support. Moreover, JACK has predefined structured patterns to manage critical situations (e.g., concurrency, reaction to events, failure). In the last years, there have been many efforts to support the IA communication in different fields, including the Industrial systems, the result has been a growing development of multi-purpose languages (e.g., 2APL, 3APL) [15]. An agent has several others features that can be considered implicitly included in the previous features (e.g., learn capability, reasoning, auto-improvement).
In complex industrial systems there are several kinds of agents according to different processes or tasks. For example, the communication-agents are specialized in communication-activities between sets of agents. They deal with strategies and protocols for the information or data exchange. Another example is the learning-agents that learn about the functionality and potentiality of the surrounding environment and transmit this knowledge to the other agents. Several others dedicated agents are developed according to specific requirements. Beyond the potentiality of a single IA the algorithms that drive the industrial processes are multi-agent based.
MASs can be considered as composed by heterogeneous agents, which are aimed to achieve common objectives. Likewise to the intelligent agent definition, also in this case there are several similar definitions of MAS [16]. In our context a MAS can be described as a suitable and reasoning assemble of agents that, according to their features, achieve common objectives through connections and teamwork intelligent mechanisms. Also the MASs have the same features observed for the intelligent agents: the main difference is the coordination and cooperation processes that drive the agents. Moreover, a MAS has interaction properties that define the collective reasoning of the whole system. Usually, each MAS has a set of rules and protocols that are inherited from each single agent. These last define the kind of collective intelligence characterizing each single MAS. These considerations conclude the paradigm definition of MASs.
A complex MAS can be composed by different sets of MASs obtaining a system with high levels of communication processes. In any case, these systems are characterized from more complex coordination, cooperation and teamwork mechanisms. Usually, no centralized control methods are implemented, but specific strict rules and hierarchic behavior strategies are performed to ensure the correct working of the system. The difficulty in the MASs management depends on the number of agents and their complexity. Figure 5 shows a typical MAS
scheme where each process is driven by a set of agents that have different dedicated assignments.
Each process (X, Y, Z and others) is composed by some tasks that are accomplished by several agents. As presented in Figure 5, the process X is composed by four industrial activities (T1, T2, T3 and T4, where T = Task), and two service activities (E and I, where E = Executor, I = Interface). All the processes involved in the industrial environment, independently from their role, are aimed to reach the common target: the objective (artifact/activity).
The agents tied to the industrial activities accomplish specific tasks tied to the industrial process, while the others two activities provide coordination and cooperation mechanisms. I-IA and E-IA work on the whole process. In fact, the first provides the result of the Industrial process (e.g., artifact or part of it, data, information) to the environment; the second uses a communication channel to interact with other processes (i.e., with the related I-IAs). Note that only the I-IA and E-IA can make actions outside their own process. The other four agents (T1-IA, T2-IA, T3-IA and T4-IA) can accomplish only internal assignments and can interact only with agents belonging to the same process.
4. MASs: Approches and Algorithms
The new generation of the industrial systems can be conceived like Distributed Artificial Intelligence (DAI) systems described as cooperative systems where a set of heterogeneous agents act jointly to solve a given problem [17]. A DAI system belongs to the Distributed Problem Solving (DPS) field where a problem is solved by using several modules [18] which cooperate in dividing and sharing knowledge about the common issue. All these factors influence the choice of the multi-agent architecture to drive the algorithms used into each agent and the related communication strategies. There are several multiagent architectures used to perform the activities of a large set of heterogeneous environments; the most effective are inspired by the distributed version of BDI (Belief-Desire-Intention) [19]. This architecture is based on the correspondence between beliefs, desires and intentions with reciprocal identifiable data structures. The BDI based systems have several features (e.g., agent scalability, real-time communication, acknowledge mechanisms). They can be managed to be adapted in any environment. In Figure 6 a general overview of the basic patterns to implement industrial architectures is given. They are based on the classical configurations used in others fields, the differences are came out during the implementation by considering the features of each agent in accordance to the specific industrial system.
As a general rule, each task is assigned to a specific Task Master Agent (TMA) which, according to the spe-
Figure 6. Sample of architectural industrial system.
cific architecture, can solve different assignments (e.g., safety and security, execution, coordination, cooperation). The TMA can work alone or supported by several subagents, depending on the nature and complexity of the task. These sub-agents deal with only a peculiar and homogeneous step of the task. Moreover, they are usually execution-oriented (i.e., executor agents) or information-oriented (i.e., acknowledger agents). Commonly, the sub-agents can have relationships only with sub-agents of the same task or with the related TMA. A particular kind of TMA, the Gateway Task Master Agent (GTMA), has to interact with all the TMAs belonging to a same process. Finally, the Process Master Agent (PMA) is used to allow interaction of different agents. Each GTMA can interact only with the related PMA, each PMA can interact with any other PMA. The communication activities are usually performed by well-known protocols and algorithms in the network field. Indeed, each communication, in the MASs, tends to be implemented according to FIPA specifications [20]. It is the IEEE Computer Society organization that promotes agent-based technology and the interoperability of its standards with other technologies.
The MASs communication activity can be analyzed by graph theory [21], where each node is an intelligent agent and the edges between nodes are the communication channels. This allows to adopt the well-known algorithms to manage environments composed by multi-agent systems. The current MASs communication processes are inspired to the basic algorithms on network architectures [22]: Routing, broadcasting, and semi-group computation.
The routing algorithms regard the processes used to visit a graph to reach a particular objective. In literature exists different algorithms to accomplish this task [23]. For example, the Dijkstra and Bellman-Ford algorithms research a shortest-path from a single source node to any other node in the graph. A variant of these algorithms is given by Floyd-Warshall and Johnson that provide strategies to research a shortest-path from multiple source nodes. Also the Prim and Kruskal algorithms (minimum spamming tree) and the Ford-Fulkerson and Karp algorithms (maximum flow) are commonly used in routing issues. All these algorithms are used to allow the agents different communication strategies which are tied to several factors, such as: type and aim of the interaction, kind of activity, kind of involved agents, and so on.
The broadcasting algorithms provide a direct way to communicate to a node in a graph. These algorithms are based on one-to-all philosophy, where a single node (i.e., agent) has to send (i.e., interact) a message to all the nodes of the graph. The weighs in the graph can drive the communication process according to the specific objective. They are usually represented by vectors that include a wide set of information (e.g., priority access value, identification value (ID), agent state value). In industrial contexts these algorithms are used to support control activities and to provide commands and scheduling steps to technical processes.
The semi-group computation algorithms are based on the binary communication of the intelligent agents composing a definite and closed set of agents. In general, the binary communication is used to reach each node in the graph through a dynamic bridge that allows two agents to communicate according to some convenience rule (e.g., distance between the agents, cost of the connection, priority level). These algorithms are usually applied, in industrial environments, where a strict intelligent agent structure is required. The particular structure of a semi-group of agent is suitable to implement systems able to react to non-deterministic events (e.g. external operations, unexpected broken events, sabotages). In addition to the three mentioned basic algorithms (and their evolutions) on network architectures, there are several experience-based approaches for everyday use which provide ad-hoc solutions allowing a qualitative improvement of the management standards. In this way, the human experience can be utilized to optimize the agent implementation improving the industrial process.
Table 1 presents a brief description of the communication algorithms according to the different main industrial environments. The physical connection of a MAS is a crucial point of the current and next generation Industrial environments, in fact the way in which sets of agents are connected each other can drive the chooses about the adopted interaction communication technologies. For example, in large-scale management processes centralized and hierarchical architectures are preferred. On the contrary, in advanced manufacturing processes are used flexible, scalable and modular architectures. Figure 7 shows that the architectures for industrial systems are