Massimo Cossentino

Multi-Agent Systems

Reinforcement learning agents can successfully learn in a variety of difficult tasks. A fundamental problem is that they may learn slowly in complex environments, inspiring the development of speedup methods such as transfer learning. Transfer improves learning by reusing learned behaviors in similar tasks, usually via an inter-task mapping, which defines how a pair of tasks are related. This paper proposes a novel transfer learning technique to autonomously construct an inter-task mapping by using a novel combinations of sparse coding, sparse projection learning, and sparse pseudo-input gaussian processes. Experiments show successful transfer of information between two very different domains: the mountain car and the pole swing-up task. This paper empirically shows that the learned inter-task mapping can be used to successfully (1) improve the performance of a learned policy on a fixed number of samples, (2) reduce the learning times needed by the algorithms to converge to a policy on a fixed number of samples, and (3) converge faster to a near-optimal policy given a large amount of samples.

Process models for agent-based development

A great deal of research in the area of agent-oriented software engineering (AOSE) focuses on proposing methodologies for agent systems, i.e., on identifying the guidelines to drive the various phases of agent-based software development and the abstractions to be exploited in these phases. However, very little attention has been paid so far to the engineering process subjacent to the development activity, disciplining the execution of the different phases involved in the software development. In this paper, we focus on process models for software development and put these in relation with current researches in AOSE. First, we introduce the key concepts and issues related to software processes and present the various software process models currently adopted in mainstream software engineering. Then, we survey the characteristics of a number of agent-oriented methodologies, as they pertain to software processes. In particular, for each methodology, we analyze which software process model it (often implicitly) underlies and which phases of the process are covered by it, thus enabling us to identify some key limitations of currently methodology-centered researches. On this basis, we eventually identify and analyze several open issues in the area of software process models for agent-based development, calling for further researches and experiences.

Agent-oriented software engineering

Considering the great number of agent-oriented methodologies that can be found in the literature, and the fact that each one defines its own concepts and system structure, one of the main challenges in agent-oriented software engineering (AOSE) research is how to make these methodologies interoperable. By defining concepts used in a specific domain in a non-ambiguous way, meta-modelling may represent a step towards such interoperability. Consequently the main objective of the AOSE TFG (Technical Forum Group) is to establish a strategy for identifying a common meta-model that could be widely adopted by the AOSE community. This paper sums up the approach used by this TFG which consists of (i) studying and comparing the meta-models related to some existing methodologies (ADELFE, Gaia, INGENIAS, PASSI, RICA and Tropos) in order to find commonalities and (ii) giving a clear and basic definition for the core concepts used in multi-agent systems for relating and positioning them in a unified MAS meta-model. The first proposal, set up by the working group, for this unified meta-model then concludes this paper.

Patterns reuse in the PASSI methodology

Design patterns already proved successful in lowering the development time and number of errors of object-oriented software; now, they are, candidate to play a similar role in the MAS (multi-agent system) context. In this work we describe our experiences in the identification, production and application of patterns for agents. Some patterns are described together with the classification criteria and documentation approach we adopt. Upon them, we base a pattern reuse process that can be considered one of the distinguishing elements of the design methodology (PASSI) we use to develop MAS. Patterns can be applied to an existing agent or used to produce a new one with the support of a specific web based application that can read both the JAVA source code and XMI representation of the agent design documentation. After the successful application of the desired pattern(s), the source code and the design diagrams (usually a structural and dynamic diagram) of the agent can be exported. Some experimental results are reported in order to demonstrate the utility of this approach in automatically producing an interesting percentage of code lines.

A possible approach to the development of robotic multi-agent systems

The design of an agent system for robotics is a problem that involves aspects coming from many different disciplines (robotics, artificial intelligence, computer vision, software engineering). The most difficult part of it, often consists in producing and tuning the algorithms that incorporates the robot behavior (planning, obstacle avoidance,...) and abilities (vision, manipulation, navigation,...). Frequently, the reuse of these parts is left to a copy and paste procedure from previous applications to the new one. In so doing many problems could arise. We propose a comprehensive approach for multi-agent systems oriented to robotics applications that uses a complete design methodology supported by a specific design tools and a pattern repository that interacting each other and with the designer allow the production of a coherent design that easily incorporates patterns coming from previously experienced features and automatically produces a large part of the final code.

Modeling ontologies for robotic environments

On the basis of a multiple abstraction levels specification process, we developed a representational model for environmental robotic knowledge through the definition of a set of ontologies using a multi perspective approach. A general ontological model for typical indoor environments has been first developed, followed by its specialization using an implementation perspective. Actual software implementation of the ontology has been obtained via a XML-based markup language, used to build a repository for robotic environmental knowledge.

Composition of a new process to meet agile needs using method engineering

The need of developing a new software engineering process to allow the quick prototyping of some robotic applications and meet the requests by some companies for a development process that was shorter than PASSI, gave us the opportunity of applying our studies on the assembling of a new SEP by reusing parts (called method fragments) from other processes. In this paper we discuss our approach that, starting from the method engineering paradigm, adapts and extends it considering specific agent-oriented issues like the multi-agent system meta-model. The final result of our experiment (Agile PASSI) is presented together with the requirements that motivated its structure.

Experiences with cicerobot, a museum guide cognitive robot

The paper describes CiceRobot, a robot based on a cognitive architecture for robot vision and action. The aim of the architecture is to integrate visual perception and actions with knowledge representation, in order to let the robot to generate a deep inner understanding of its environment. The principled integration of perception, action and of symbolic knowledge is based on the introduction of an intermediate representation based on Gardenfors conceptual spaces. The architecture has been tested on a RWI B21 autonomous robot on tasks related with guided tours in the Archaeological Museum of Agrigento. Experimental results are presented.

Handbook on Agent-Oriented Design Processes

To deal with the flexible architectures and evolving functionalities of complex modern systems, the agent metaphor and agent-based computing are often the most appropriate software design approach. As a result, a broad range of special-purpose design processes has been developed in the last several years to tackle the challenges of these specific application domains. In this context, in early 2012 the IEEE-FIPA Design Process Documentation Template SC0097B was defined, which facilitates the representation of design processes and method fragments through the use of standardized templates, thus supporting the creation of easily sharable repositories and facilitating the composition of new design processes. Following this standardization approach, this book gathers the documentations of some of the best-known agent-oriented design processes. After an introductory section, describing the goal of the book and the existing IEEE FIPA standard for design process documentation, thirteen processes (including the widely known Open UP, the de facto standard in object-oriented software engineering) are documented by their original creators or other well-known scientists working in the field. As a result, this is the first work to adopt a standard, unified descriptive approach for documenting different processes, making it much easier to study the individual processes, to rigorously compare them, and to apply them in industrial projects. While there are a few books on the market describing the individual agent-oriented design processes, none of them presents all the processes, let alone in the same format. With this handbook, for the first time, researchers as well as professional software developers looking for an overview as well as for detailed and standardized descriptions of design processes will find a comprehensive presentation of the most important agent-oriented design processes, which will be an invaluable resource when developing solutions in various application areas.

Engineering Multi-Agent Systems

Smart spaces are inherently complex and dynamic systems, where diverse devices, sensors, actuators and computational elements need to interact with one another. A middleware infrastructure can provide suitable abstractions that simplify the task, and allow designers to ignore the details of the underlying elements. Unfortunately, however, existing middleware solutions do not generalize well to different kinds of spaces, since they often fail to address the scalability and dynamism of such spaces. In this paper we propose SeSaMe, a semantic and self-adaptive middleware infrastructure for highly dynamic and massive smart spaces. SeSaMe establishes a “backbone” to let components connect to the system without any prior knowledge of its topology. It is capable of maintaining the system’s overall reliability, even when multiple components leave or fail unexpectedly, and of coping with message congestion, by dynamically altering the system’s topology. SeSaMe also provides a simple declarative language for defining how one wants the system to evolve over time, and semantic technologies for harmonizing the interaction of different kinds of components. The main new features of SeSaMe are exemplified on two example smart spaces with significantly different characteristics.