Partially observable stochastic games (POSGs) are a Markovian formalism used to model a set of agents acting in a stochastic environment, in which each agent has its own reward function. As is common with multi-agent decision making problems, the model and runtime complexity is exponential in the number of agents, which can be prohibitively large. Lifting is a technique that treats groups of indistinguishable instances through representatives if possible, yielding tractable inference in the number of objects in a model. This paper applies lifting to the agent set in POSGs, yielding so-called isomorphic POSGs that have a model complexity no longer dependent on the number of agents, and presents a lifted solution approach that exploits this lifted agent set for space and runtime gains.
Multi-label classification (MLC) is a supervised learning problem where each instance can be associated with none, one, or multiple labels. MLC has received increasing attention due to its wide range of applications, such as text categorization and medical diagnosis. Despite a rich literature on MLC, handling imbalanced data, often encountered in real-world MLC datasets, has not been tackled satisfactorily. Based on a thorough literature review, it appears that the existing methods for imbalanced MLC are either hard to be coupled with sound theoretical guarantees or of limited scalability. This paper discusses the potential (dis)advantages of existing methods for imbalanced MLC, when being coupled with Binary relevance classifier (BRC), and introduces Discrete Minimax BRC (DMBRC), which would be a promising attempt to robustify the BRC by leveraging theoretically sound properties of the Discrete Minimax Classifier. We also provide empirical evidence to illustrate how DMBRC may be advantageous in balancing the label-wise error rates. Finally, we envision future works on further strengthening DMBRC in both label-wise error rates and conventional MLC evaluation metrics.
Recently, it was shown that a radial basis function network (RBFN) with a softmax output layer amounts to pooling by Dempster's rule positive and negative evidence for each class, and approximating the resulting belief function by a probability distribution using the plausibility transform. This so-called latent belief function offers a richer uncertainty quantification than the probabilistic output of the RBFN. In this paper, we show that there exists actually a set of latent belief functions for a RBFN. This set is obtained by considering all possible dependence structures, which are described by correlations, between the positive and negative evidence for each class. Furthermore, we show that performance can be enhanced by optimizing the correlations brought to light.
This short paper presents a description of SDF-FuzzIA, a Fuzzy-Ontology LLM-based system for the intelligent analysis of geothematic data that serves as a plug-in to the Sustainability Decision Framework (SDF) Decision Support System (DSS). A description of the components implemented in the system is given, followed by an explanation of the interaction between the components and the main system. As this still is a work in progress, future directions and possible hurdles are explored.
The development of secure and efficient solutions for access control is an important issue in a variety of applications. One of the main challenges is to avoid situations that make access control decision-making impossible. However, avoiding such situations hampers the evolution of the model, as it means either adding a large set of constraints or dealing with each conflict situation. It is, therefore, important to explore methods that deal with conflicts as they arise while providing explanations of the decision taken. In this work, we develop an ontology to manage dynamic and abstract access control rules based on the OrBAC (Organization-Based Access Control) model and integrate an ordering relation over instances of the ontology. Our method takes advantage of the application of inconsistency-tolerant semantics to resolve conflicts and generate explanations for transparency in decisions made. Our results show that the approach efficiently preserves the consistency of the decision taken and provides potentially useful and human-friendly explanations.
Clustering is an unsupervised machine learning task that aims to discover natural groups in the given dataset. K-mode algorithms, which are adaptions of K-mean algorithms for continuous data, are among the most popular algorithms for discovering clusters in categorical data. In this paper, we present some results on how to accelerate them using the triangle inequality, while still always computing exactly the same result as the original K-mode algorithms. We also provide some empirical evidence to illustrate the potential gains provided by leveraging the triangle inequality. Finally, we envision future work aimed at providing a comprehensive understanding of the use of triangle inequality in accelerating other clustering algorithms for categorical data.
In the framework of Qualitative Spatio-Temporal Reasoning (QSTR), we can consider constraints like x { is above ∨ is under } y, and combinations thereof, to represent and reason about spatial or temporal information in an intuitive, human-like way. QSTR becomes particularly important in view of possible lack, uncertainty, and/or imperfection of metric data, as treating such quantitative information qualitatively would provide more leeway to perform sound reasoning. Adding to the usefulness of QSTR, in this short paper, we introduce the notion of multi-scenario for representing solutions of networks of qualitative spatio-temporal constraints in a compact manner, as a means to assessing and enhancing the explainability and robustness of AI systems that involve spatio-temporal information. Further, we prove certain theoretical properties pertaining to this novel notion, and even introduce some robustness measures relating to our notion of multi-scenario.
Provenance calculus is based on two operations forming a semi-ring. Its aim is to combine and propagate annotations associated with data. This note emphasizes the point that this calculation, if based on max and min operations, exactly corresponds to query evaluation when data are labeled with levels of certainty (necessity) in the sense of possibility theory.
We introduce Bel coalitional games, that generalize classical coalitional games, where uncertainty is modelled through the Dempster-Shafer theory and every agent can have different knowledge. We propose the notion of contract in our framework, that specifies how agents divide the values of the coalitions and we use the Choquet integral to model the agents' preferences between contracts. Next, we study the core under two different moments of the game by defining the ex-ante core and the ex-t-interim core, where, in the latter, we need the Dempster conditional rule to update the mass functions of agents. In particular, in the last step of the ex-t-interim case and when the set of states reduces to singleton, i.e. when there is no uncertainty, we recover the classical definition of the core. Finally, we show some results about the ex-ante and the ext-interim core of Bel coalitional games, following the well-known results about classical coalitional games.
Dialogue game-based semantics for abstract argumentation are relevant for several reasons. From a theoretical point of view, they provide a different perspective regarding extension-based or labeling-based approaches to study the theoretical properties of the argument evaluation process. From a more practical perspective, they allow us to examine whether or not an argument belongs to a given extension (or labeling) without computing an entire (set of) extension(s) or labeling(s), and guide the development of efficient algorithms. This last point is significant in the context of the development of argumentation-based knowledge representation and reasoning tools for real-world applications. In this paper, we expand the dialogue game-based semantics available for extended abstract argumentation frameworks, a generalization of abstract argumentation frameworks where two kinds of defeat are considered, proper and blocking, and the sub-argument relation is taken into account. The novel dialogue game-based semantics we propose is inspired in a specific interpretation of cycles that considers them as an indecision, a situation in which we do not have enough information to decide the status of the arguments involved.
For static lifted inference algorithms, completeness, i.e., domain liftability, is extensively studied. However, so far no domain liftability results for temporal lifted inference algorithms exist. In this paper, we contribute the first completeness and complexity analysis for a temporal lifted algorithm, the so-called lifted dynamic junction tree algorithm (LDJT), which is the only exact lifted temporal inference algorithm out there. To handle temporal aspects efficiently, LDJT uses conditional independences to proceed in time, leading to restrictions w.r.t. elimination orders. We show that these restrictions influence the domain liftability results and show that one particular case while proceeding in time, has to be excluded from FO2. Additionally, for the complexity of LDJT, we prove that the lifted width is in even more cases smaller than the corresponding treewidth in comparison to static inference.
Global Sensitivity Analysis aims at explaining how much each random variable contributes to the variance of the output of a black-box model. The standard approach – namely Sobol indices – computes the contribution of each subset of variables but requires that the variables are independent. The Shapley effect (based on the Shapley value) has been defined for dependent variables, but gives the contribution of each variable individually instead of the contribution of subsets of variables. The aim of this work is to propose a novel approach for dependent variables that defines the level of contribution of each subset of variables so that they sum up to the total variance of the output of the model. We show that we come up with known concepts – namely the Banzhaf values and interaction indices, up to a multiplicative factor.
Recent advances in deep learning have often surpassed human performance in image classification. Among the most renowned cases, just think of the ImageNet Large Scale Visual Recognition Challenge competition. However, challenges persist in complex fields such as medical imaging. An example is the Human Protein Atlas which maps all human proteins in more than 171,000 images that makes a computation challenge due to high class imbalance. To address these challenges from a green perspective, we propose a transfer learning approach using Convolutional Neural Networks (CNNs) pre-trained on the ImageNet dataset. We use CNN layers as feature extractors, feeding the extracted features into a Support Vector Machine with a linear kernel. Our method combines both image-level and cell-level perspectives. Furthermore, at the cell level, we segment nuclei and extract the surrounding nuclear membrane area. The combination of the two perspectives shows promising classification performance with limited computational effort.
Existing works on stability of incomplete argumentation frameworks (IAFs) discuss the status of an argument or a set of arguments under a given semantics during the changing of an IAF towards complete AFs. We argue that the stability of an IAF itself is worth studying, i.e., checking whether all extensions under a semantics are the same in every completion of the IAF. When an IAF becomes stable in this sense, there is no need to investigate the uncertain arguments or attacks within the IAF as in the end all its complete AFs will share the same extensions. We further present a relaxed notion called weak stability so that the same extensions are required solely within the shared arguments of every two completions rather than all arguments in the IAF. In this paper we study six common semantics of AF and show precise complexity results of checking the (weak) stability of an IAF, which turns to be more complex than checking a set of arguments whereas less complex than checking an argument. Meanwhile, we give a SAT encoding for the stability problems with coNP-c complexity.
In this paper, we present a reinforcement learning approach for resolving inconsistencies in qualitative constraint networks ( s). s are typically used in constraint programming to represent and reason about intuitive spatial or temporal relations like x {is inside of ∨ overlaps } y. Naturally, s are not immune to uncertainty, noise, or imperfect data that may be present in information, and thus, more often than not, they are hampered by inconsistencies. We propose a multi-armed bandit approach that defines a well-suited ordering of constraints for finding a maximal satisfiable subset of them. Specifically, our learning approach interacts with a solver, and after each trial a reward is returned to measure the performance of the selected action (constraint addition). The reward function is based on the reduction of the solution space of a consistent reconstruction of the input . Early experimental results obtained by our algorithm suggest that we can do better than the state of the art in terms of both effectiveness, viz., lower number of repairs obtained for an inconsistent , and efficiency, viz., faster runtime.
We investigate the problem of measuring inconsistency in linear temporal logic on finite traces (LTLf). In particular, we present Answer Set Programming-based approaches to compute a selection of traditional inconsistency measures w.r.t. LTLf knowledge bases. In contrast to existing works (mostly on propositional logic), these approaches are novel in the sense that they allow to assess logical inconsistency in presence of temporal operators, as offered by LTLf. In an experimental evaluation on real-world data from the area of business process management, we show that our approaches are practically feasible.
The increase in computing power in recent years has brought generative models and the use of synthetic data back to the fore to solve a variety of previously unsolved problems, in particular when fields are subject to constraints linked to the sensitivity of the information processed. This article proposes a modified version of restricted Boltzmann machines (RBM), known as Bernoulli machines, to improve its ability to handle non-binary data without making the methodology more complex to understand and manipulate. To assess the performance of our algorithm, we compare it with various generative models that are well documented and have repeatedly proven their effectiveness in a variety of contexts. We also chose to use a large number of open source datasets with different types of features and different sizes in order the verify the generalization capacity and sclalability of our approach.
In this paper, we investigate the problem of eliciting information from an expert, where the assumed uncertainty model is a coherent upper prevision (or equivalently a closed convex set of probabilities). The goal is to solve a decision problem under the maximality decision rule, with as few queries to the expert as possible. To address this, we study the range of coherent upper bounds an expert may give on a given query. In doing so, we provide new results and characterisations for this range. We then use these results to provide an algorithm of elicitation. We illustrate the algorithm on an example.
The Gödel integral is a variant of the Sugeno integral. It is an expressive aggregation operator to compute a global evaluation combining local values taken for a set of considered criteria and a capacity, a set function that represents the importance of these criteria and their interactions. A crucial question is then the elicitation of this capacity so as to match the user preferences. Such a task is common in many fields, such as XAI, where counterfactual explanations can be generated in the form of counterfactual examples by aggregating a pair of objective and subjective criteria. In this paper, we focus on approaches to elicit such a capacity when the user preferences are expressed as an ordering relation between classes of equivalent objects. We first propose a theoretical characterisation of the set of admissible capacities: it allows to determine lower and upper bounds of possible capacities when it is not empty and it provides explanations when it is empty. Then, we introduce the algorithm GICEP (for Gödel Integral for Capacity Elicitation from Preference relations), to compute these bounds and to provide explanations. Experiments conducted on synthetic data show the relevance and efficiency of GICEP.
We consider the problem of learning one of three possible fuzzy generalizations of the Jaccard similarity measure, based on the d-Choquet integral. Each of the resulting fuzzy similarity measures is parameterized by a capacity and by a real parameter. The capacity describes the weights assigned to groups of attributes and their interactions, while the real parameter is related to the restricted dissimilarity function used to evaluate differences among attributes. To face identifiability issues and in view of an XAI use of the learned capacity, the parameters’ set is restricted to the set of (at most) 2-additive completely monotone capacities. Next, under a suitable definition of entropy for completely monotone capacities, we address different entropic regularization schemes to single out interactions between groups of attributes. This is done by taking as reference a local uniform Möbius inverse over sets of attributes with the same cardinality.
