Catálogo de publicaciones - libros

In this paper, we present a model that allows agents to reason about adjusting their autonomy in multiagent systems, integrating both full transfers of decision making control to other entities (users or agents) and initiations of interaction to gather more information (referred to as partial transfers of control). We show how agents can determine the optimal transfer of control strategy (which specifies which entities to transfer control to, and how long to wait for a response), by generating and evaluating possible transfer of control strategies. This approach extends earlier efforts in the field by explicitly demonstrating how information seeking interaction can be integrated into the overall processing of the agent. Through examples, we demonstrate the benefits of an agent asking questions, in order to determine the most useful transfers, or to improve its own decision making ability. In particular, we show how the model can be used to effectively determine whether or not it is beneficial to initiate interaction with users. We conclude with discussions on the value of the model as the basis for designing adjustable autonomy systems.

Computing optimal coalition structures is an important research problem in multi-agent systems. It has rich application in real world problems, including logistics and supply chains. We study computing optimal coalition structures in linear production domains. The common goal of the agents is to maximize the system’s profit. Agents perform two steps: ) deliberate profitable coalitions, and ) exchange computed coalitions and generate coalition structures. In our previous studies, agents keep growing their coalitions from the singleton ones in the deliberation step. This work takes opposite approach that agents keep pruning unlikely profitable coalitions from the grand coalition. It also relaxes the strict condition of coalition center, which yields the minimal cost to the coalition. Here, agents merely keep generating profitable coalitions. Furthermore, we introduce new concepts, i.e., and , in our algorithm and provide an example of how it can work. Lastly, we show that our algorithm outperforms exhaustive search in generating optimal coalition structures in terms of elapsed time and number of coalition structures generated.

Assistance to people suffering from cognitive deficiencies in a smart home raises complex issues. Plan recognition is one of them. We propose a formal framework for the recognition process based on lattice theory and action description logic. The framework minimizes the uncertainty about the prediction of the observed agent’s behaviour by dynamically generating new implicit extra-plans. This approach offers an effective solution to actual plan recognition problem in a smart home, in order to provide assistance to persons suffering from cognitive deficits.

Thanks to new technological advances, geospatial information is getting easier to disseminate via Internet and to access using mobile devices. Currently, several mapping applications are providing thousands of users worldwide with web and mobile maps generated automatically by extracting and displaying pre-processed data which is stored beforehand in specific databases. Though rapid, this approach lacks flexibility. To enhance this flexibility, the mapping application must determine by itself the spatial information that should be considered as relevant with respect to the map context of use. It must also determine and apply the relevant transformations to spatial information, autonomously and on-the-fly, in order to adapt it to the user’s needs. In order to support this reasoning process, several knowledge-based approaches have been proposed. However, they did not often result in satisfactory results. In this paper, we propose a multiagent-based approach to improve real-time web and mobile map generation in terms of personalization, data generation and transfer. To this end, the agents of our system compete for space occupation until they are able to generate the required map. These agents, which are assigned to spatial objects, generate and transfer the final data to the user simultaneously, in real-time.

We are interested in contributing to solving effectively a particular type of real-time stochastic resource allocation problem. Firstly, one distinction is that certain tasks may create other tasks. Then, positive and negative interactions among the resources are considered, in achieving the tasks, in order to obtain and maintain an efficient coordination. A standard Multiagent Markov Decision Process (MMDP) approach is too prohibitive to solve this type of problem in real-time. To address this complex resource management problem, the merging of an approach which considers the complexity associated to a high number of different resource types (i.e. Multiagent Task Associated Markov Decision Processes (MTAMDP)), with an approach which considers the complexity associated to the creation of task by other tasks (i.e. Acyclic Decomposition) is proposed. The combination of these two approaches produces a near-optimal solution in much less time than a standard MMDP approach.

So far, most equilibrium concepts in game theory require that the rewards and actions of the other agents are known and/or observed by all agents. However, in real life problems, agents are generally faced with situations where they only have partial or no knowledge about their environment and the other agents evolving in it. In this context, all an agent can do is reasoning about its own payoffs and consequently, cannot rely on classical equilibria through deliberation, which requires full knowledge and observability of the other agents. To palliate to this difficulty, we introduce the satisfaction principle from which an equilibrium can arise as the result of the agents’ individual learning experiences. We define such an equilibrium and then we present different algorithms that can be used to reach it. Finally, we present experimental results that show that using learning strategies based on this specific equilibrium, agents will generally coordinate themselves on a Pareto-optimal joint strategy, that is not always a Nash equilibrium, even though each agent is individually rational, in the sense that they try to maximize their own satisfaction.

Our literature review revealed that several applications successfully simulate certain kinds of human behaviors in spatial environments, but they have some limitations related to the ‘believability’ and the ‘usability’ of the simulations. This paper aims to present a set of requirements for multiagentbased simulations in terms of ‘believability’ and ‘usability’. It also presents how these requirements have been put into use to develop a multiagent-based simulation prototype of customers’ shopping behavior in a mall. Using software agents equipped with spatial and cognitive capabilities, this prototype can be considered sufficiently ‘believable’ and ‘usable’ for end-users, mainly mall managers in our case. We show how shopping behavior simulator can support the decision-making process with respect to the spatial configuration of the shopping mall.

We focus on clustering gene expression temporal profiles, and propose a novel, simple algorithm that is powerful enough to find an efficient distribution of genes over clusters. We also introduce a variant of a clustering index that can effectively decide upon the optimal number of clusters for a given dataset. The clustering method is based on a profile-alignment approach, which minimizes the mean-square-error of the first order differentials, to hierarchically cluster microarray time-series data. The effectiveness of our algorithm has been tested on datasets drawn from standard experiments, showing that our approach can effectively cluster the datasets based on profile similarity.

Gliomas are diffuse, invasive brain tumors. We propose a 3D classification-based diffusion model, CDM, that predicts how a glioma will grow at a voxel-level, on the basis of features specific to the patient, properties of the tumor, and attributes of that voxel. We use Supervised Learning algorithms to learn this general model, by observing the growth patterns of gliomas from other patients. Our empirical results on clinical data demonstrate that our learned CDM model can, in most cases, predict glioma growth more effectively than two standard models: uniform radial growth across all tissue types, and another that assumes faster diffusion in white matter.

There are some neural network applications in proteomics; however, design and use of a neural network depends on the nature of the problem and the dataset studied. Bayesian framework is a consistent learning paradigm for a feed-forward neural network to infer knowledge from experimental data. Bayesian regularization automates the process of learning by pruning the unnecessary weights of a feed-forward neural network, a technique of which has been shown in this paper and applied to detect the glycosylation sites in epidermal growth factor-like repeat proteins involving in cancer as a case study. After applying the Bayesian framework, the number of network parameters decreased by 47.62%. The model performance comparing to One Step Secant method increased more than 34.92%. Bayesian learning produced more consistent outcomes than one step secant method did; however, it is computationally complex and slow, and the role of prior knowledge and its correlation with model selection should be further studied.