Jason induction of logical decision trees: a learning library and its application to commitment

  • Authors:

  • Alejandro Guerra-Hernández;Carlos Alberto González-Alarcón;Amal El Fallah Seghrouchni

  • Affiliations:

  • Departamento de Inteligencia Artificial, Universidad Veracruzana, Facultad de Física e Inteligencia Artificial, Xalapa, Ver., México;Departamento de Inteligencia Artificial, Universidad Veracruzana, Facultad de Física e Inteligencia Artificial, Xalapa, Ver., México;Laboratoire d'Informatique de Paris 6, Université Pierre et Marie Curie, Paris, France

  • Venue:
  • MICAI'10 Proceedings of the 9th Mexican international conference on Advances in artificial intelligence: Part I
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper presents JILDT (Jason Induction of Logical Decision Trees), a library that defines two learning agent classes for Jason, the well known java-based implementation of AgentSpeak(L). Agents defined as instances of JILDT can learn about their reasons to adopt intentions performing first-order induction of decision trees. A set of plans and actions are defined in the library for collecting training examples of executed intentions, labeling them as succeeded or failed executions, computing the target language for the induction, and using the induced trees to modify accordingly the plans of the learning agents. The library is tested studying commitment: A simple problem in a world of blocks is used to compare the behavior of a default Jason agent that does not reconsider his intentions, unless they fail; a learning agent that reconsiders when to adopt intentions by experience; and a single-minded agent that also drops intentions when this is rational. Results are very promissory for both, justifying a formal theory of single-mind commitment based on learning, as well as enhancing the adopted inductive process.