Henry Soldano

Learning in BDI multi-agent systems

This paper deals with the issue of learning in multi-agent systems (MAS). Particularly, we are interested in BDI (Belief, Desire, Intention) agents. Despite the relevance of the BDI model of rational agency, little work has been done to deal with its two main limitations: i) The lack of learning competences; and ii) The lack of explicit multi-agent functionality. From the multi-agent learning perspective, we propose a BDI agent architecture extended with learning competences for MAS context. Induction of Logical Decision Trees, a first order method, is used to enable agents to learn when their plans are successfully executable. Our implementation enables multiple agents executed as parallel functions in a single Lisp image. In addition, our approach maintains consistency between learning and the theory of practical reasoning.

ZooM: a nested Galois lattices-based system for conceptual clustering

This paper deals with the representation of multi-valued data by clustering them in a small number of classes organized in a hierarchy and described at an appropriate level of abstraction. The contribution of this paper is three fold. First, we investigate a partial order, namely nesting, relating Galois lattices. A nested Galois lattice is obtained by reducing (through projections) the original lattice. As a consequence it makes coarser the equivalence relations defined on extents and intents. Second we investigate the intensional and extensional aspects of the languages used in our system ZooM. In particular we discuss the notion of α-extension of terms of a class language £. We also present our most expressive language £3, close to a description logic, and which expresses optionality or/and multi-valuation of attributes. Finally, the nesting order between the Galois lattices corresponding to various languages and extensions is exploited in the interactive system ZooM. Typically a ZooM session starts from a propositional language £2 and a coarse view of the data (through α-extension). Then the user selects two ordered nodes in the lattice and ZooM constructs a fine-grained lattice between the antecedents of these nodes. So the general purpose of ZooM is to give a general view of concepts addressing a large data set, then focussing on part of this coarse taxonomy.

SMILE: Sound Multi-agent Incremental LEarning

This article deals with the problem of collaborative learning in a multi-agent system. Here each agent can update incrementally its beliefs B (the concept representation) so that it is in a way kept consistent with the whole set of information K (the examples) that he has received from the environment or other agents. We extend this notion of consistency (or soundness) to the whole MAS and discuss how to obtain that, at any moment, a same consistent concept representation is present in each agent. The corresponding protocol is applied to supervised concept learning. The resulting method SMILE (standing for Sound Multi-agent Incremental LEarning) is described and experimented here. Surprisingly some difficult boolean formulas are better learned, given the same learning set, by a Multi agent system than by a single agent.

Alpha galois lattices: an overview

What we propose here is to reduce the size of Galois lattices still conserving their formal structure and exhaustivity. For that purpose we use a preliminary partition of the instance set, representing the association of a “type” to each instance. By redefining the notion of extent of a term in order to cope, to a certain degree (denoted as α), with this partition, we define a particular family of Galois lattices denoted as Alpha Galois lattices. We also discuss the related implication rules defined as inclusion of such α-extents and show that Iceberg concept lattices are Alpha Galois lattices where the partition is reduced to one single class.

Learning better together

This article addresses collaborative concept learning in a MAS. In a concept learning problem an agent incrementally revises a hypothetical representation of some target concept to keep it consistent with the whole set of examples that it receives from the environment or from other agents. In the program SMILE, this notion of consistency was extended to a group of agents. A surprising experimental result of that work was that a group of agents learns better the difficult boolean problems, than a unique agent receiving the same examples. The first purpose of the present paper is to propose some explanation about such unexpected superiority of collaborative learning. Furthermore, when considering large societies of agents, using pure sequential protocols is unrrealistic. The second and main purpose of this paper is thus to propose and experiment broadcast protocols for collaborative learning.

Abstract concept lattices

We present a view of abstraction based on a structure preserving reduction of the Galois connection between a language L of terms and the powerset of a set of instances O. Such a relation is materialized as an extension-intension lattice, namely a concept lattice when L is the powerset of a set P of attributes. We define and characterize an abstraction A as some part of either the language or the powerset of O, defined in such a way that the extension-intension latticial structure is preserved. Such a structure is denoted for short as an abstract lattice. We discuss the extensional abstract lattices obtained by so reducing the powerset of O, together together with the corresponding abstract implications, and discuss alpha lattices as particular abstract lattices. Finally we give formal framework allowing to define a generalized abstract lattice whose language is made of terms mixing abstract and non abstract conjunctions of properties.

Incremental Learning of Relational Action Rules

In the Relational Reinforcement learning framework, we propose an algorithm that learns an action model allowing to predict the resulting state of each action in any given situation. The system incrementally learns a set of first order rules: each time an example contradicting the current model (a counter-example) is encountered, the model is revised to preserve coherence and completeness, by using data-driven generalization and specialization mechanisms. The system is proved to converge by storing counter-examples only, and experiments on RRL benchmarks demonstrate its good performance w.r.t state of the art RRL systems.

Active learning of relational action models

We consider an agent which learns a relational action model in order to be able to predict the effects of his actions. The model consists of a set of STRIPS-like rules, i.e. rules predicting what has changed in the current state when applying a given action as far as a set of preconditions is satisfied by the current state. Here several rules can be associated to a given action, therefore allowing to model conditional effects. Learning is online, as examples result from actions performed by the agent, and incremental, as the current action model is revised each time it is contradicted by unexpected effects resulting from his actions. The form of the model allows using it as an input of standard planners. In this work, the learning unit IRALe is embedded in an integrated system able to i) learn an action model ii) select its actions iii) plan to reach a goal. The agent uses the current action model to perform active learning, i.e. to select actions with the purpose of reaching states that will enforce a revision of the model, and uses its planning abilities to have a realistic evaluation of the accuracy of the model.

Incremental construction of alpha lattices and association rules

In this paper we discuss Alpha Galois lattices (Alpha lattices for short) and the corresponding association rules. An alpha lattice is coarser than the related concept lattice and so contains fewer nodes, so fewer closed patterns, and a smaller basis of association rules. Coarseness depends on a a priori classification, i.e. a cover C of the powerset of the instance set I, and on a granularity parameter α. In this paper, we define and experiment a Merge operator that when applied to two Alpha lattices G(C1, α) and G(C2, α) generates the Alpha lattice G(C1∪C2, α), so leading to a class-incremental construction of Alpha lattices. We then briefly discuss the implementation of the incremental process and describe the min-max bases of association rules extracted from Alpha lattices.

Distributed learning in intentional BDI multi-agent systems

Despite the relevance of the belief-desire-intention (BDI) model of rational agency, little work has been done to deal with its two main limitations: the lack of learning competences and explicit multi-agent functionality. Our work deals with the problem of designing BDI learning agents situated in a multi-agent system (MAS). From the MAS learning perspective, we have proposed an extended BDI architecture, where agents are able to perform induction of first-order logical decision trees. These agents learn about their practical reasons to adopt a plan as an intention. Particularly, induction is used to update these reasons (the context of plans), if a plan fails when executed, after it had been selected to form an intention. Here, we emphasize the way MAS concepts, as cooperative goal adoption, enable distributed forms of learning, e.g., distributed data gathering. Consistency between learning and the theory of practical reasoning is guaranteed, i.e., learning is just another competence of the agents, performed under BDI rationality.