Dr. Alejandro Guerra-Hernández | Investigación
Learning under Rational Agency
Among the rational models of agency, the BDI intentional approach is very well known in the agent community. It is grounded on two philosophical concepts: Intentional systems à la Dennett; and Michael Bratman's theory of practical reasoning. The model is formally well founded in a serie of multi-modal logics, widely studied in the 90's. Different BDI architectures have been implemented, e.g., PRS, dMARS, !jam. Some succesful applications of the model are reported, e.g., monitoring the space shuttle. On the other hand, two limitations of the model were well known: The lack of social, multi-agent explicit functionality; and the lack of learning competences.
Amal El Fallah-Seghrouchni proposed me a PhD thesis (Guerra2003) on the problem of learning in a Multi-Agents System (MAS) composed by intentional rational agents, i.e., the limitations of the BDI model! We were working with Henry Soldano at Paris 13, adopting the induction of logical decision trees to allow our BDI agents to learn about the "reasons" they have to form an intention at run-time. These agents can do it all alone, or communicating with other agents in a MAS (Guerra2001,2004a,2004b,2005a).
From the learning perspective, it seems that we need some incremental version of the first-order inductive algorithm for decision trees, we used. Wulfrano Ramírez implement an ID4 like algorithm for introducing incrementally in Tilde. Henry Soldano proposed us a new protocol for the distributed MAS learning, it was tested using propositional representations, so that we are interested on going first-order (it seems the incremental algorithm is required here). Gustavo Ortíz (Guerra2008a), reimplemented our approach using Jason, an interpreter of AgentSpeak(L) and defined an Smile-like protocol, although it was less distributed than Smile given the nature of Tilde and ID3 (Guerra2008b).
Recently we were interested in approaching intention revision from the intentional learning perspective. In order to do that, we need to know the kind of commitment strategy followed by AgentSpeak(L) agents. Martin Castro-Manzano and me, we have proposed CTL AgentSpeak(L) as a formalism for verifying such properties (Guerra2009,2008d). A proposal of this subject, although less technical, has been reported earlier (Guerra2008c).
Currently, Carlos González-Alarcón is working on the project 4. He has already implement the link between ACE/Tilde, an external inducer of logical decision trees, and a class of agents defined in Jason. Now, he's implementing in Java the induction algorithm to get the full Jason's learning library.
Projects: 1) Propose an AgentSpeak(L) full theory of sigle-mind commitment and verify it using CTL AgentSpeak(L). 2) Compare AgentSpeak(L) and Geogeff's system I using the logics BDI CTL and CTL AgentSpeak(L). 3) Implement a full incremental ID5 like algorithm for induction of first-order decision trees. 4) Implement a Learning "standard" library for Jason, the Java implementation of an extended AgentSpeak(L) interpreter.
Intelligent learning interfaces
An interface agent is a software component complementary to a user interface, in which agent and user engage in a colaborative process which includes communication, event monitoring and task ejecution. The agent’s world is the interface which in turn is the user-agent interaction medium. The agent acts independently of the interface in benefit of the user´s goals and without his explicit intervention.
The metaphore for this kind of interaction is that of a personal assistant. A personal assistant must have a great deal of autonomy so that it can act independently in the ejecution of specific tasks, without constant external intervention or supervision. It must be robust so that it is less vulnerable to failure, and it should have some knowledge about the user’s habits and preferences. Since, in general, the user’s behavior changes in time, so that it is not always possible to predict his future actions by means of fixed rules, the agent must exhibit some kind of adaptive functionality in order to be useful.
I worked with Ana Silvia Agüera and Manuel Martinez (Agüera2000) exploring the possibility of applying Memory Based Reasoning (MBR) to adaptive behavior in interface agents. This research topic is related to Pattie Maes inteface agents, it was my first reading on autonomous agents, and inspired me to do my thesis on action selection mechanisms, that was a more natural first step. Anyway, Ana Silvia did her thesis on learning interface agents, under de direction of Manuel and my not always opportunistic help.
Lately we were interested in offering support to experts in the domain of data mining and knowledge discovery in data bases. Nicandro Cruz Ramírez and I propose a multi-agent system inmersed in WEKA to explore such approach (Guerra2008e). Rosibelda Mondragón did her master's thesis in this subject. Sonia Mestizo implemented a help center based on an expert system for the web, the idea is now to use multi-agent systems instead of the expert system to gain proactivity.
Currently Omar Chiyéan is working on project 2. He has already implemented a set of primitive actions that enable an agent to extract the bibliography from a PDF document using drag-and-drop. He has also test some actions to exploit the Google API and recover the PDF of all available references in de original document and to maintain a BIbtex DB of the recovered references. The idea now is to use such DB to create a social based, semantic web service based on a MAS. For instance, the agents of co-referenced members of a research team emerged in the MAS and share the bibliography for local searches. Or a report of productivy of the research team can be generated for each agent in the MAS via co-references too. It's all about exploring social uses of such knowledge.
Projects: 1) Although we know how to situate AgentSpeak(L) agents in WEKA, more interesting experiments are required to understand the benefits of such approach. The subject is designing a set of such experiments and justify the use of a multi-agent system in data mining. 2) Exploring the use of MAS in the searching, organization and use of bilbiography; particularily under a task oriented perspective.
Action selection mechanisms
One of the key problems in modeling and designing autonomous agents is how choice of alternative actions should be performed. The action selection problem is defined as follows: given a set of actions that an agent can perform, given a set of goals that the agent must archive, given perceptual sequences from environment, how can the agent decide its next action in order to progress toward goals archivement? Three different aspects of this question have been identified:
- The chosen action should be appropriate to the agent’s internal and external circumstances, that is, goals and environment;
- Conflicts between different actions should be solved in such a way that mutually contradictory actions are not executed at the same time, and switching between actions is effected smoothly and coherently;
- The behavior of the agent should serve goal archivement, including of course, homeostatic goals.
There are two fundamental questions from the designer perspective:
- Should action selection be a centralized or a decentralized process?
- Should conflict resolution use a fixed priority or should it be dynamically reconfigurable?
Ourselves and others, argue that in order to gain robustness, action selection must be decentralized, and that autonomy requires dynamically reconfigurable conflict solution. We have studied the behavior network mechanism, originally proposed by Pattie Maes, and proposed different tools and extensions (Guerra1997,1997b,1998). These tools have been used to model the behavior of the howler monkey Allouata Palliata, in the Monots Project. The main tool we have used is [ABC], a behavior network simulator implented in Tk/Tcl 8.0. The ziped file include the slides of a tutorial [pdf] about this action selection mechanism.
Our most recent work on ABC (Guerra2005) extends this tool to study plasticity in behavior networks based agents, using reinforcement learning. The case study is again the foraging of howler monkeys (Selene Ivette Jiménez master's thesis 2005).