Communications of the ACM
Taming Heterogeneous Agent Architectures
The recent advances in network-based software applications and the advent of ubiquitous computing are pushing us inevitably toward a world of autonomous software architectures. This trend has spurred the revitalization of agent technology as a complement to the object paradigm for a variety of modern application domains, including e-commerce, software development environments, and personal assistants. There is explicit evidence indicating the penetration of software agents is also high for systems used in military and government contexts [2, 8]. Agents, like objects, provide services to their clients, but are recognizably different from objects as seen from an architectural point of view [3, 5, 8]. Unlike objects, an agent is an autonomous entity that takes the initiative to achieve system goals and represent software users [2, 3].
As a result, architects of software agents are faced with basic concerns, such as the agent services that are made available to the clients, and a number of additional concerns on top of the basic concerns. The internal architecture of a single agent encompasses multiple properties, including autonomy, interaction, adaptation, collaboration, learning, and mobility. Hence, architects of agent-based systems are also concerned with issues such as making an agent interact appropriately, handling the agent's adaptive behavior, structuring the agent's autonomous behavior, designing the agent roles and protocols for inter-agent collaboration purposes, and incorporating learning mechanisms into the agent's structure in a modular manner.
Not surprisingly, separation of concerns is at the core of the development of agent-based software systems [5, 12]. The reuse and maintenance of agent elements depend largely on the ability of used architectural abstractions to support the separate handling of agent-specific concerns since an early state of design. The applied architectural styles must enable the modularization of each agent concern and their proper composition, so that the achieved segregation significantly limits the impact of a change and improves the chances for architecture reuse in other software projects. This separation of concerns needs to be guaranteed throughout the different development phases, especially from the architectural to the implementation phase. The architectural separation will be lost if the implementation abstractions are not able to preserve it. In fact, the sole use of existing well-known agent platforms, such as JADE and JACK [12], do not provide advanced mechanisms to achieve separation of concerns, which are restricted to exploit the object-oriented composition and decomposition mechanisms [5]. On the other hand, even though such advanced implementation mechanisms are delivered, the benefits of separation of agent concerns would be hindered if existing architectural abstractions do not allow the achievement of proper system modularization from the design outset.
Heterogeneity in Agent Architectures
Although separation of concerns is critical to architects of agent-based software, it is often difficult to achieve in realistic systems for several reasons. First, these systems typically encompass heterogeneous types of agents [2]. The internal architecture of distinct agent types differs widely from each other since they incorporate distinct properties [8]. Second, each property is orthogonal and interacts with the agent's basic functionality and often with other agent properties. These properties typically crosscut several modules of an agent architecture, independently from the adopted internal model, such as the constraint-oriented model or the BDI (Belief-Desire-Intention) model [10], and from the agent's cognitive level, such as reactive agents, deliberative agents, or even hybrid agents. Third, these crosscutting agent-specific concerns are related in dramatically different ways that depend on the agents' types and the application requirements [6]. For example, in a given agent-based application, the mobility behavior of an agent may only directly affect the basic functionality of the agent, while it may also crosscut the collaboration and interaction concerns in a second application.
As a consequence, agent-based applications require an architectural approach that is flexible enough to support adjustable composition of agent concerns and the construction of heterogeneous agent architectures according to the application demands. This flexibility requirement is even more stringent in open agent-based systems due to their adaptive and open nature. In these contexts, the agent architecture needs to be modular enough to support the dynamic reconfiguration of its internal elements. For example, roles potentially need to be changed as the agent moves to new environments. The degree of autonomy and the learning strategies may also need to be adapted or disabled according to the dynamic execution contexts. While part of the system-level reconfiguration facilities are typically provided by specific middleware implementations, many of the agent concerns and their composition are essentially application-dependent. Software architects need to prepare and conceive at the design outset modular agent architectures in order to cope with these heterogeneity issues.
Agent-based applications require an architectural approach that is flexible enough to support adjustable composition of agent concerns and the construction of heterogeneous agent architectures according to the application demands.
Shortcomings of Classical Agent Architectures
Agent-oriented software engineering has been studied from different perspectives, including agent-oriented methodologies and languages for higher-level development phases, conceptual modeling, and implementation frameworks [2, 3]. Although separation of concerns is widely recognized as crucial to the development of maintainable multi-agent software [5, 6, 12], existing approaches do not scale up to support the separation of agent properties in heterogeneous architectures. Developers must rely on traditional architectural patterns, such as the Mediator pattern [4] and the Layers pattern [4], in order to build their systems. As illustrated in Figure 1, these solutions define architectural abstractions, such as mediators and layers, and composition rules to support the isolation of agent concerns and their further composition.
However, these solutions impose rigid connections on the architectural components, which make the construction of heterogeneous agent types difficult and not scalable to cope with the complexity of multiple interactive agent concerns. These existing architectural proposals are applicable to simple agent architectures and with few agenthood properties. The constraints on the composition rules imposed by those solutions do not scale up to master the intrinsic crosscutting character of agent properties in heterogeneous contexts. Figure 1 illustrates this problem in terms of a layered agent architecture (Figure 1a) and a mediator-based agent architecture (Figure 1b), represented with a simplified UML notation.
The layered architecture imposes a bidirectional communication only between adjacent layers; that is, the architectural components of a software agent. In order to make proper internal decisions and adapt the agent knowledge accordingly, the autonomy and adaptation components must be aware of received messages and external stimulus coming from the surrounding environment. As illustrated in Figure 1a, this requirement forces the software architects to add interaction-related interfaces into the other layers so that the autonomy and adaptation components have access to the perceived stimulus and messages. As a result, the interaction concern (represented in blue) crosscuts the modularity of the kernel and adaptation components.
The situation is not different in the mediator-based solution. This architectural style requires all the inter-component communications being intermediated by a central component, which is also in charge of encapsulating the agent's basic functionality. As illustrated in Figure 1b, this architectural approach also fails to isolate the interaction concern and leads to the intermingling of the agent kernel functionality and the message-processing functionality. Note that a similar problem occurs with the mobility concern as the communication with the autonomy component needs to be settled by the agent kernel. Hence, the inability of traditional architectural styles and respective composition rules in capturing the crosscutting feature of agent concerns in heterogeneous architectures leads to several negative consequences, such as architectural tangling and bloated interfaces, which are described in detail in the table here.
Spreading the Crosscutting to Subsequent Artifacts
The problems associated with heterogeneous agent architectures are typically disseminated from the agent architecture specifications to artifacts generated later in the software life cycle. Hence, the crosscutting of agent concerns is inevitably spanned over the resulting artifacts in the detailed design level and implementation. It does not matter what kind of decomposition and abstractions the agent-based software developers are relying on; the problem is the same if you are using an agent-oriented design language [2], or an OO modeling language, such as UML.
For example, the architectural abstractions and composition rules are not directly supported in OO design and programming languages. They are not aligned with the composition and decomposition mechanisms of the object paradigm, which makes it difficult to handle the heterogeneous agent types in a single system [6]. For instance, the inheritance mechanism usually leads to large inheritance trees with replication of code and an explosion of the number of classes [5, 6, 12]. Even in some simple agent architectures, bringing them into the detailed design and code also raises similar problems (examples are shown in the table), such as scattering and tangling.
Figure 2a illustrates how the reification of heterogeneous agent architectures in the design tends to be scattered over many classes of the system design. It shows a partial representation of an agent-based system [6] composed of three agent types, each with a different internal architecture. Each set of classes, surrounded by a gray rectangle, has the main purpose of modularizing a specific agent concern, namely the agents' basic functionality and collaboration. The figure shows that agent-specific concerns crosscut the classes implementing those surrounded classes, such as "Information Agent" and "Role." The use of OO mechanisms breaks not only the class-level modularity, but also the operation-level modularity. Figure 2b shows the partial implementation of a searcher agent, which consists of a class and respective operations with a confusing tangle of lines of code for different concerns. On the left, there is the code for a class as a non-agent entity. On the right, there is an equivalent class with code for implementing agent-specific concerns. After analyzing the code on the right, it is clear that such code has lost the functional encapsulation of the non-agent version (left side). Moreover, it is a confusing intermingling of lines of code for different agent concerns, and the modularization of several agent-specific concerns is lost across the system classes.
Architecting Software Agents with Aspects
Aspect-oriented software development [9] is an evolving paradigm to modularize concerns that existing software engineering abstractions and mechanisms are not able to capture explicitly. The notion of aspects encourages modular descriptions of complex software by providing support through new composition mechanisms for cleanly separating the system functionality from its crosscutting concerns. However, existing aspect-oriented approaches have not been explored in the context of heterogeneous agent architectures; they have focused on the context of classical crosscutting concerns, such as distribution [11], persistence [11], and design patterns [7]. These concerns have been mostly studied in an isolated way and from the implementation point of view.
Aspects and its new composition possibilities can be exploited at the architectural level to capture the multiple interacting agent concerns that are difficult to modularize with existing architectural abstractions. An aspect-oriented architectural style brings to the software architects a new abstraction, the notion of architectural aspects, and new composition means for handling each crosscutting agent concern as an individual component at an early stage of design, as illustrated in Figure 3a. Each architectural aspect modularizes a typical crosscutting agent property and separates it from the agent kernel. The "aspectization" of agent architectures allows the association of agenthood properties with the basic functionality in a way that is transparent to the agent kernel.
The key idea to enable adjustable compositions is the notion of crosscutting interfaces, which are modularity abstractions attached to the architectural aspects. A crosscutting interface is different from a module interface in the sense that the latter essentially provides services to other components. Crosscutting interfaces provide services to the system, but also specify when and how an aspect affects other architectural components. Contrary to interfaces in traditional architecture styles, they flexibly determine which external components and interfaces the architectural aspect of a software agent will be connected. With this dependency inversion, crosscutting interfaces overcome the problems associated with the rigidness implicit in traditional architectural styles (see Figure 1). Each agent's architectural aspect can be more flexibly composed with the agent kernel and with any agent aspects depending on the requirements of a specific agent architecture.
Each of the architectural aspects is related to more than one component, representing the crosscutting nature of agent properties in complex architectures. An agent's architectural aspect can realize more than one crosscutting interface since it can crosscut multiple agent components in different ways. The interface of an architectural aspect can crosscut the Kernel component and other architectural aspects. An aspect interface crosscuts either the internals of an agent component or elements of other interfaces. The first case means the architectural aspect affects the internal structure or dynamic behavior of an agent component. The second case means the aspect affects elements of an interface.
Instances of aspect-oriented agent architectures are illustrated for a user agent (see Figure 3a) and an information agent (see Figure 3b). Extensions of UML are used to represent the architectural aspects (components with diamonds on their top), crosscutting relationships (dotted arrow), and crosscutting interfaces (small diamonds). Their partial non-aspect-oriented detailed designs were represented in Figure 2a. The user agent is prominently reactive, while the information agent is a proactive entity. As a result, their architectures are very different. For example, the component in charge of implementing the autonomous behavior has different interfaces and relationships with each component in the agent architecture. Reactive agents make decisions only as a matter of deciding about external requests embedded in incoming messages. Proactive agents also need to inspect changes in their internal state and decide for starting new actions by their own initiative independently from requests from other agents and system users. In other words, the autonomy component also needs to observe events in the agent kernel and in the agent roles (collaboration component) in order to make decisions.
Figure 3 shows that the aspect-oriented style easily supports both architectural configurations. A partial architectural design is shown and some non-aspectual components, such as the ones for agent knowledge representation and message assembling, are left out for simplicity purposes. In the first case (Figure 3a), the autonomy aspect does not have an interface for enabling proactive behavior. It has two crosscutting interfaces that affect the kernel and interaction components. In the second case (Figure 3b), the autonomy component includes a third crosscutting interface that supports the observation of the agent kernel and the collaboration component in order to trigger the proactive behavior of the information agent. The architectures have other important differences in their configurations and compositions. For example, the architecture of the information agent also includes the mobility and learning aspects and a third interface to enable sensory behavior. Moreover, the adaptation aspect crosscuts the sensory interface in order to adapt as new environmentally relevant events are sensed by the agent.
What are Other Benefits?
The benefits of having new architectural design rules entailed by an aspect-oriented style are not limited to improved composability of agent architecture concerns. We have noticed a number of other positive consequences such as the ones discussed here.
Modular Reasoning at the Architectural Level. The aspectization of agent architectures supports modular reasoning, since the architects are able to more independently treat each agenthood property. Different from other architectural styles, the agent concerns are modularized and do not affect the definitions and interfaces of multiple architectural modules. The notion of crosscutting interfaces allows addressing tangling- and scattering-related problems typically found in the definition of heterogeneous agent architectures. As a result, software architects are able to make decisions about an agent property while only looking at its description and its interface with other concerns.
Smooth Transition in Software Life-Cycle Phases. Aspect-oriented agent architectures are directly mapped to implementation abstractions using well-known aspect-oriented programming languages, such as AspectJ [1]. AspectJ, which extends the Java programming language, is the most popular aspect language. Architectural aspects are decomposed into a set of AspectJ aspects and classes. The crosscutting interfaces are realized as pointcuts, advice, and inter-type declarations, because they define different ways an aspect affects other design and implementation modules. Join points are well-defined points in the dynamic execution of the system components. Examples of join points are method calls and method executions. Pointcuts have a name and are collections of join points. Advice is a special method-like construct attached to pointcuts. Inter-type declarations introduce attributes, methods, and interface implementation declarations into the components to which the crosscutting interface is attached. We have implemented an AspectJ framework that supports this aspect-oriented architectural style at the implementation level [5]. It helps to guarantee a smooth transition from the specification of heterogeneous agent architectures to their detailed design and implementation. We have also developed several other aspect-oriented techniques to cope with agent aspects in different software development phases in order to facilitate the traceability of the software engineering artifacts.
Coping with Dynamic Adaptability and Customizability. Architecting software agents with improved separation of concerns is of paramount importance in open, dynamic agent-based systems. Dynamic reconfiguration of agent roles and collaboration protocols are often required as the agents move to different environments. In addition, with the growing number of applications for pervasive computing, the selection of learning and coordination strategies may depend on the context in which the agent is being executed. As such, agent architectures need to be designed properly, and the dependency inversion gained with aspect-oriented agent architectures is a key factor to allow the agent adaptation and customization.
Conclusion
Numerous types of agent architectures are prevalent [8] and must be developed in a way that meets the stringent modern requirements of evolvability, reusability, and dynamic reconfigurability. Existing architectural styles are rigid by their very nature and unable to cope with the crosscutting nature of agent properties as well as the complexity of heterogeneous agent architectures, which are often required in realistic modern systems. On the other hand, aspect-oriented software development is gaining wide attention both in research environments and in industry as a paradigm to promote improved modularity of complex software systems. The exploration of aspect-oriented techniques clearly seems a promising step forward to allow the construction of more flexible agent architectures and to foster enhanced quality of realistic multi-agent systems.
References
1. AspectJ Web site. The AspectJ Guide; eclipse.org/aspectj/.
2. Bergenti, F., Gleizes, M.-P., and Zambonelli, F., Eds. Methodologies and Software Engineering for Agent Systems: The Agent-Oriented Software Engineering Handbook, Volume 11. Springer-Verlag, 2004.
3. Bordini, R., Dastani, M., Dix, J., and Seghrouchni, A., Eds. Multi-Agent Programming: Languages, Platforms, and Applications. Springer, NY, 2005.
4. Buschmann, F. et al. Pattern-Oriented Software Architecture: A System of Patterns. Wiley, NY, 1996.
5. Garcia, A., Lucena, C., and Cowan, D. Agents in object-oriented software engineering. Software: Practice and Experience 34, 5 (Apr. 2004), 489521.
6. Garcia, A. et al. Separation of concerns in multi-agent systems: An empirical study. In C. Lucena et al., Eds., Software Engineering for Multi-Agent Systems II, Springer-Verlag, LNCS 2940 (Mar. 2004), 4972.
7. Garcia, A. et al. Modularizing design patterns with aspects: A quantitative study. In Proceedings of the 4th ACM International Conference on Aspect-Oriented Software Development (AOSD'05), 111.
8. Jennings, N. An agent-based approach for building complex software systems. Commun. ACM 44, 4 (Apr. 2001), 3541.
9. Kiczales, G. et al. Aspect-oriented programming. In Proceedings of the European Conference on Object-Oriented Programming-ECOOP'97, LNCS 1241, 220242.
10. Rao, A. and Georgeff, M. BDI agents: From theory to practice. In V. Lesser and L. Gasser, Eds., Proceedings of the 1st International Conference on Multi-Agent Systems (ICMAS-95), 312319.
11. Soares, S, Laureano, E., and Borba, P. Implementing distribution and persistence aspects with AspectJ. In Proceedings of the ACM Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA'02), 174190.
12. Ubayashi, N. and Tamai, T. Separation of concerns in mobile agent applications. In Proceedings of the 3rd International Conference Reflection 2001, LNCS 2192, 89109.
Figures
Figure 1. Crosscutting concerns in heterogeneous agent architectures.
Figure 2. Heterogeneous agent architectures: Crosscutting the life-cycle artifacts.
Figure 3. Aspectizing heterogeneous agent architectures: Enabling flexible composition.
©2008 ACM 0001-0782/08/0500 $5.00
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2008 ACM, Inc.