Despite the plethora of frameworks and tools for developing agent systems, there is a remarkable lack of generalised methodologies for assessing their performance. Existing methods in the field of software engineering do not adequately address the unpredictable and complex nature of intelligent agents. We introduce a generic methodology for evaluating agent performance; the Agent Performance Evaluation (APE) methodology consists of representation tools, guidelines and techniques for organizing, categorizing and using metrics, measurements and aggregated characterizations of agent performance. The core of APE is the Metrics Representation Tree, a generic structure that enables efficient manipulation of evaluation-specific information. This paper provides a formal specification of the proposed methodology in Z notation and demonstrates how to apply it on an existing multi-agent system.
Storytelling is to generate a logical narrative that connects a sequence of events in the story. While the narrative coherence is the most important element in the storytelling, the diversity prolongs the story life and elicits interests on interpreting the story. In this paper, we investigate the narrative diversity given that the storytelling process is modelled by Bayesian networks. Bayesian networks structure causal relations between variables using a graphical representation, which facilitates communication between story generators and readers. The storytelling is equivalent to propagating the events in Bayesian networks, which preserves the narrative coherence. By adding the sampling process in the propagation, we can see the emergence of the narrative diversity in the storytelling. We study the entire process for a plot in one classical Chinese tale.
It is quite complicated to develop a multi-agent simulation system. In this paper, we realize a simulation case in a simple way that just chooses some components and combines them; while other groups of people are responsible for developing these components. The “accessories library” consists of these relative independent functional components which are abstracted from various simulation cases. Therefore, we propose the “accessories library” and a platform based on multi-roles, which helps domain experts find their roles and create complex multi-agent systems easily. We further describe the process of conducting every role involved. In order to confirm the role-based platform and the respective processes, we take a football field case as an example to illustrate the implementation.
We tackle the issue of modeling the temporal propagation dynamics in multiplex networks where agents are connected by multiple linking types. As information can be disseminated across multiple layers, cross-layer delay is one of the main aspects of temporal dynamics apart from the intra-layer delay between agents in each layer. First, the independent cascade model is generalized to multiplex networks incorporating the two aspects of temporal dynamics. Then, numerical modeling based on the mean-filed assumption is proposed to grasp the macroscopic behaviors of temporal propagation dynamics in multiplex networks, and the microcosmic influence of the two aspects of temporal dynamics is discussed with the aid of two simple case studies. By contrast to the multi-agent simulation, numerical modeling can predict the propagation in multiplex networks by a slight margin. Moreover, we show that the two aspects of temporal dynamics can dramatically vary the influence of multiple layers on information spreading in multiplex networks.
We present in this paper a reputation model for agent-based web services grouped into communities by their equivalent functionalities. The reputation of each web service is based on the non-functional properties of its interactions with other web services from the same community. We exploit various clustering and anomaly detection techniques to analyze and identify the quality patterns provided by each service. This enables the master of each community to allocate the requests it receives to the web service that best fulfill the quality requirements of the service consumers. Our experiments present realistic scenarios based on synthetic data that characterizes the reputation feedback of the quality provided by a web service at different times. The results showcase the capability of our reputation model in portraying the quality of web services that reside in a community and characterizing their fair and unfair feedback reports.
Support vector machine is a supervised learning model with associated learning algorithms that analyzes data and recognizes patterns. In various applications, the SVM shows its advantage of the classification performance, however, the original SVM was designed for the numerical data. For using the SVM on the nominal data, most previous research used a certain number to replace each nominal value or transformed the nominal value into the one hot vector. Both methods could not present the original nominal data’s structure and the similarity between them, which leads to information loss from the data and reduce the classification performance. In this work, we design a novel coupled similarity metric between nominally attributed data. This metric is pairwise, we also propose an adapted SVM which can handle this. The experiment result shows the proposed method outperforms the traditional SVM and other popular classification methods on various public data sets.
In this paper, we present an agent-based solution of meta-learning problem which focuses on optimization of data mining processes. We exploit the framework of computational multi-agent systems in which various meta-learning problems have been already studied, e.g. parameter-space search or simple method recommendation. In this paper, we examine the effect of data preprocessing for machine learning problems. We perform the set of experiments in the search-space of data mining processes which is constituted by combinations of preprocessing methods with classifiers. The optimization takes place with respect to two criteria - error-rate and model learning time, which are partially complementary. The results of the consistent search algorithm on a number of classification data-sets are shown and the advantage of automated preprocessing augmentation of method recommendation is demonstrated.
Communication is one of central activities in multiagent systems. It enables the knowledge sharing among multiple agents and improves the planning quality in a long run. In this paper, we study communication decision problems in the framework of interactive dynamic influence diagrams~(I-DIDs). I-DIDs are recognized probabilistic graphical models for sequential decision making in uncertain multiagent settings. We extend the representation to explicitly model communication actions as well as their relations to other variables in the domain. The challenging work is on developing an incentive mechanism that drives level 0 agents to learn communication while they act alone in a dynamic environment. We present solutions to the new model and show meaningful communication strategies in a multiagent problem domain.
Advanced recommender systems of the third generation (3G) emphasize employment of semantically clear models of customer cross-domain profile learned using all available data sources. The paper focuses on conceptual level of ontology-based formal model of the customer profile built in actionable form. Learning of cross-domain customer profile as well as its use in recommendation scenario requires solving a number of novel problems, e.g. information fusion and data source privacy preservation, among others. The paper proposes an ontology-driven personalized customer profile model and outlines an agent-based architecture supporting implementation of interaction-intensive agent collaboration in two variants of target decision making procedure that are content-based and collaborative filtering both exploiting semantic similarity measures.
With the rapid proliferation of social media and online community, a vast amount of text data has been generated. Discovering the insightful value of the text data has increased its importance, a variety of text mining and process algorithms have been created in the recent years such as classification, clustering, similarity comparison. Most previous research uses a vector-space model for text representation and analysis. However, the vector-space model does not utilise the information about the relationships between the term to term. Moreover, the classic classification methods also ignore the relationships between each text document to another. In other word, the traditional text mining techniques assume the relation between terms and between documents are independent and identically distributed (iid). In this paper, we will introduce a novel term representation by involving the coupled relations from term to term. This coupled representation provides much richer information that enables us to create a coupled similarity metric for measuring document similarity, and a coupled document similarity based K-Nearest centroid classifier will be applied to the classification task. Experiments verify the proposed approach outperforming the classic vector-space based classifier, and show potential advantages and richness in exploring the other text mining tasks.
This paper presents a new approach to modelling the time course of user-experience (UX). Flexibility in modelling is essential: to select or develop UX models based on the outcome variables that are of interest in terms of explanation or prediction. At the same time, there is potential for (partial) re-using UX models across products and generalisation of models. As a case study, an experience model is developed for a particular consumer product, based on a time-sequential framework of subjective well-being [13] and a theoretical framework of flow for human-computer interaction [23]. The model is represented as a dynamic Bayesian network and the feasibility and limitations of using DBN are assessed. Future work will empirically evaluate the model with users of consumer products.
Feature sets in many domains often contain many irrelevant and redundant features, both of which have a negative effect on the performance and complexity of agents that use the data [9]. Supervised feature selection aims to overcome this problem by selecting features that are highly related to the class labels, yet unrelated to each other. One proposed technique to select good features with few inter-dependencies is minimal Redundancy Maximal Relevance (mRMR) [12], but this can be impractical with large feature sets. In many situations, features are extracted from signal data such as vehicle telemetry, medical sensors, or financial time-series, and it is possible for feature redundancies to exist both between features extracted from the same signal (intra-signal), and between features extracted from different signals (inter-signal). We propose a two stage selection process to take advantage of these different types of redundancy, considering intra-signal and inter-signal redundancies separately. We illustrate the process on vehicle telemetry signal data collected in a driver distraction monitoring project. We evaluate it using several machine learning algorithms: Random Forest; Naïve Bayes; and C4.5 Decision Tree. Our results show that this two stage process significantly reduces the computation required because of inter-dependency calculations, while having minimal detrimental effect on the performance of the feature sets produced.
In most information retrieval systems, software processes reason about passive data. Our approach instantiates each piece of information as an agent that actively seeks to organize itself with respect to other agents (including queries). Imitating the movement of bodies under physical forces, we describe a distributed algorithm ("force-based clustering," or FBC) for dynamically clustering and querying large, heterogeneous, dynamic collections of entities. The algorithm moves records in a virtual space in a way that estimates the transitive closure of the pairwise comparisons. We demonstrate FBC on a large, heterogeneous collection of records, each representing a person. We have some information about a person of interest, but no record in the collection directly matches this information. Application of FBC identifies a small subset of records that are good candidates for describing the per-son of interest, for further manual investigation and verification.
Extracting norms from computer-mediated human interactions is gaining popularity since huge volume of data is available from which norms can be extracted. Open source communities offer exciting new application opportunities for extracting norms since such communities involve developers from different geographical regions, background and cultures. Investigating the types of norms that exist in open source projects and their efficacy (i.e. the usage of norms) in enabling smoother functioning however has not received much attention from the normative multi-agent systems (NorMAS) community. This paper makes two contributions in this regard. First, it presents norm compliance results from a case study involving three open source Java projects. Second, it presents an architecture for mining norms from open source projects. It also discusses the opportunities presented by the domain of software repositories for the study of norms. In particular, it points towards how norms can be mined by leveraging and extending prior work in the areas of Normative Multi-Agent Systems (NorMAS) and mining software repositories.
Most of the clustering algorithms were designed to cluster the data in convex spherical sample space, but their ability was poor for clustering more complex structures. In the past few years, several spectral clustering algorithms were proposed to cluster arbitrarily shaped data in various real applications including image processing and web analysis. However, most of these algorithms were based on k-means, which is a randomized algorithm and makes the algorithm easy to fall into local optimal solutions. Hierarchical method could handle the local optimum well because it organizes data into different groups at different levels. In this paper, we propose a novel clustering algorithm called spectral clustering algorithm based on hierarchical clustering (SCHC), which combines the advantages of hierarchical clustering and spectral clustering algorithms to avoid the local optimum issues. The experiments on both synthetic data sets and real data sets show that SCHC outperforms other six popular clustering algorithms. The method is simple but is shown to be efficient in clustering both convex shaped data and arbitrarily shaped data.
Monitoring the electrodermal activity is increasingly accomplished in agent-based experimental settings as the skin is believed to be the only organ to react only to the sympathetic nervous system. This physiological signal has the potential to reveal paths that lead to excitement, attention, arousal and anxiety. However, electrodermal analysis has been driven by simple feature-extraction, instead of using expressive models that consider a more flexible behavior of the signal for improved emotion recognition. This paper proposes a novel approach centered on sequential patterns to classify the signal into a set of key emotional states. The approach combines SAX for pre-processing the signal and hidden Markov models. This approach was tested over a collected sample of signals using Affectiva-QSensor. An extensive human-to-human and human-to-robot experimental setting is under development for further validation and characterization of emotion-centered patterns.
Smart space is a distributed ambient environment with existing, inside it, dynamic set of inhabitants (living and nonliving) solving various own and common tasks. The mission of smart space is to provide, for its inhabitants, with context–dependent information, communication, services, reminders and personalized recommendations in a user–friendly mode where and when needed in ubiquitous and unobtrusive style. The smart space R&D uses large diversity of models, frameworks, and technologies and their integration is the first challenging smart space problem. Another challenge is caused by the necessity to process huge volumes of heterogeneous information perceived by distributed sensors in adaptive, self–organizing, learnable, and efficient style. The paper analyses these challenges and emphasizes an important role of the technology integrating agent and data mining to overcome both these challenges.
Trading agents are very useful for supporting investors in making decisions in financial markets, but the existing trading agent research focuses on simulation on artificial data. This leads to limitations in its usefulness. As for investors, how trading agents help them manipulate their assets according to their risk appetite and thus obtain a higher return is a big issue. Portfolio optimization is an approach used by many researchers to resolve this issue, but the focus is mainly on developing more accurate mathematical estimation methods, and overlooks an important factor: trading strategy. Since the global financial crisis added uncertainty to financial markets, there is an increasing demand for trading agents to be more active in providing trading strategies that will better capture trading opportunities. In this paper, we propose a new approach, namely trading strategy based portfolio selection, by which trading agents combine assets and their corresponding trading strategies to construct new portfolios, following which, trading agents can help investors to obtain the optimal weights for their portfolios according to their risk appetite. We use historical data to test our approach, the results show that it can help investors make more profit according to their risk tolerance by selecting the best portfolio in real financial markets.
The emergence of Multi-Agent systems as a software paradigm that most suitably fits all types of problems and architectures is already experiencing significant revisions. A more consistent approach on agent programming, and the adoption of Software Engineering standards has indicated the pros and cons of Agent Technology and has limited the scope of the, once considered, programming ‘panacea’. Nowadays, the most active area of agent development is by far that of intelligent agent systems, where learning, adaptation, and knowledge extraction are at the core of the related research effort. Discussing knowledge extraction, data mining, once infamous for its application on bank processing and intelligence agencies, has become an unmatched enabling technology for intelligent systems. Naturally enough, a fruitful synergy of the aforementioned technologies has already been proposed that would combine the benefits of both worlds and would offer computer scientists with new tools in their effort to build more sophisticated software systems. Current work discusses Agent Academy, an agent toolkit that supports: a) rapid agent application development and, b) dynamic incorporation of knowledge extracted by the use of data mining techniques into agent behaviors in an as much untroubled manner as possible.
When a new agent enters to an open multiagent system, bootstrapping its trust becomes a challenge because of the lack of any direct or reputational evidence. To get around this problem, existing approaches assume the same a priori trust for all newcomers. However, assuming the same a priori trust for all agents may lead to other problems like whitewashing. In this paper, we leverage graph mining and knowledge representation to estimate a priori trust for agents. For this purpose, our approach first discovers significant patterns that may be used to characterise trustworthy and untrustworthy agents. Then, these patterns are used as features to train a regression model to estimate trustworthiness. Lastly, a priori trust for newcomers are estimated using the discovered features based on the trained model. Through extensive simulations, we have showed that the proposed approach significantly outperforms existing approaches.
