CILIOS: Connectionist inductive learning and inter-ontology similarities for recommending information agents

Abstract

For a software information agent, operating on behalf of a human owner and belonging to a community of agents, the choice of communicating or not with another agent becomes a decision to take, since communication generally implies a cost. Since these agents often operate as recommender systems, on the basis of dynamic recognition of their human owners’ behaviour and by generally using hybrid machine learning techniques, three main necessities arise in their design, namely (i) providing the agent with an internal representation of both interests and behaviour of its owner, usually called ontology; (ii) detecting inter-ontology properties that can help an agent to choose the most promising agents to be contacted for knowledge-sharing purposes; (iii) semi-automatically constructing the agent ontology, by simply observing the behaviour of the user supported by the agent, leaving to the user only the task of defining concepts and categories of interest. We present a complete MAS architecture, called connectionist learning and inter-ontology similarities (CILIOS), for supporting agent mutual monitoring, trying to cover all the issues above. CILIOS exploits an ontology model able to represent concepts, concept collections, functions and causal implications among events in a multi-agent environment; moreover, it uses a mechanism capable of inducing logical rules representing agent behaviour in the ontology by means of a connectionist ontology representation, based on neural-symbolic networks, i.e., networks whose input and output nodes are associated with logic variables.

Introduction

The mutual monitoring among agents is a key issue in developing multi-agent systems (MASs) for supporting Web applications. As pointed out in [1], in multi-agent interactions, agents often share a common goal, evaluated through a global utility function. However, the agent cannot observe the global state of the environment in which it lives, therefore agents have to communicate with each other in order to share the information needed for deciding which actions to perform. Thus, for the agent, choosing whether to communicate or not with another agent becomes a decision to take, since communication generally implies a cost. In this sense, the mutual monitoring is a way for supporting cooperation among agents, although cooperation is a very broad issue, certainly not limited to mutual monitoring, and usually concerned not only with information sharing (or sharing mutual models including beliefs, desires, intentions) but particularly with goal sharing and task distribution in a joint effort of the agents. In this paper we focus on the particular issue of mutual monitoring and on its contribution in realizing an effective agent cooperation. In our opinion, three main problems arise in the scenario described above:

• An agent-based framework addressing the issue of agents mutual monitoring generally provides each agent with an internal representation of both interests and behaviour of the associated human user, usually called ontology [2], [3].

• The next step for implementing mutual monitoring is to detect inter-ontology properties that can support an agent to choose the most promising agents to be contacted for knowledge-sharing purposes. Some of these approaches use as inter-ontology properties the similarity between ontology concepts [4], also by determining synonymies and homonymies [2] between concepts, while other approaches exploit, in addition to similarity, also other properties involving information about the whole agent community, as the reputation that an agent has gained inside the community [5]. All these approaches use a symbolic¹ representation of agent ontologies. It is worth pointing out that, in order to detect inter ontology properties, some approaches assume an individualistic view of agents societies, where each agent does not have access to the ontologies of the other agents. This is the viewpoint of most of the so-called BDI approaches [6], [7]. Differently, other approaches, as [8], [9], adopt a “social” view of agent communities, where it is assumed that the ontology of each agent is, even partially, accessible for each other agent. The latter viewpoint is the one adopted in this paper where we deal with inter-ontology similarities.

• Another important issue, well known in the agent community, is the necessity that the agents automatically construct their ontologies, by observing the behaviour of the monitored users. While the interaction between an agent and its user is always necessary in defining user interests and representing them into an ontology, the agent should be capable of automatically extracting logical rules representing user behaviour and/or causal implications among events. This latter issue needs the exploitation of some inductive methods, which can be either symbolic, as in the case of the inductive logic programming, or connectionist, like neural networks and genetic algorithms. Note that, although the presence of an automatic learning mechanism is often a condition necessary to realize an effective agent autonomy, such a condition is not sufficient, as it is not the only property ensuring autonomy. For instance, an agent represented as an object on receiving a message cannot refuse to respond (it cannot autonomously choose whether to respond or not); in general, no client–server architecture may show autonomy of the server. In particular, autonomous “planning” or “scheduling of tasks” by the agent should also be guaranteed, based on the run time (learning) knowledge the agent has at its disposal.

To the best of our knowledge, all the approaches existing in the literature deal only with one or at most two of the three issues above. In this work, we present a mechanism dealing with all the three issues described above, capable of inducing logical rules representing agent behaviour in the ontology by means of a connectionist ontology representation, based on neural-symbolic networks. This mechanism exploits a new ontology representation, called information agent connectionist ontology model (IACOM). The idea underlying IACOM is that of combining some recent proposals for representing symbolic knowledge by a connectionist system, as those presented in [10], [11], with the classical use of neural-networks as function approximators. IACOM derives from [10] the idea of using a neural-symbolic network for representing a logic program. Moreover, IACOM introduces the new notion of action for combining this neural-symbolic network with classical neural networks that represent the functions contained in the ontology. In our framework, an action is defined as a mathematical relation between an event $e$ and a function $f$, and represents the fact that $f$ is activated (i.e., it is called by the system for generating an output) only if $e$ happens. This notion is implemented by means of special arcs linking the network associated with the logic program and the networks associated with the functions. The resulting network is able to represent the overall behaviour of the agent. Fig. 1 graphically describes this idea. The white nodes $e1$, $e2$ and $c$ represent three events (i.e., boolean variables) composing a neural-symbolic network, whose structure means that the event $c$, being associated with the output node, is a consequence of $e1$ and $e2$. Therefore, $c$ becomes true only if both $e1$ and $e2$ are $\mathit{true}$. The black nodes compose a traditional three layers neural network, which represents a real function. The input nodes are associated with the inputs of the function (i.e., real variables) while the output node represents the output of the function. The arc between the event $c$ and the function output is an action arc. It represents the fact that the function output is computed only if the event $c$ is true. Otherwise, if $c$ is false, the function output has a undefined value, meaning that the function in this case does not yield any output. The use of the approach presented in [10] gives, on one hand, the possibility to represent an initial background knowledge by a neural-symbolic network. Such a network can be trained for refining the initial knowledge by means of a supervised learning phase that exploits, as training set, the actual user's behaviour. On the other hand, the choice of this approach allows the obtained knowledge, represented by the network weights, to be re-translated into the symbolic form for making it understandable. Finally, the so obtained symbolic knowledge can be object of a reasoning phase that generates useful deductions.

Note that this approach of constructing an agent ontology can be considered semi-automatic, in the sense that the human user has already the necessity to define concepts and categories of interests, but the agent equipped with its learning algorithm has the capability of discovering logic rules, functions and actions involving these concepts and categories. We also present a complete architecture of a MAS, called connectionist learning and inter-ontology similarities (CILIOS), for supporting agent mutual monitoring, attempting to integrate solutions for each of the above mentioned problems. This architecture is conceived as extension of the well-known JADE platform, and can be used for implementing an agent platform for supporting many important tasks as, for instance, generating recommendations in Web communities or supporting Web navigation and file sharing. Finally, we present a series of experimental tests we have conducted for evaluating the capability of our proposed architecture in supporting agent mutual monitoring, in the case of collaborative filtering recommender systems. The paper is structured as follows: in Section 2, some related work is discussed. In Section 3 the CILIOS architecture is described in detail, while in Section 4 we give a formal description of the IAOM symbolic ontology model. In Section 5 the framework for extracting inter-ontology similarities is presented. In Section 6, we introduce the IACOM connectionist ontology model and propose a method for building agent ontologies by using neural-symbolic networks. Some experiments using CILIOS are described in Section 7 and, finally, in Section 8 we draw our conclusions, by discussing both advantages and limitations of our proposal.