Título: I-DYNA: An Architecture for Integrating Reacting, Planning, Learning and Self-Awareness in Autonomous Agents
Autores: Voinea, Camelia Florela
Resumo: This work represents a connectionist approach to adaptive reasoning. It proposes a new version of the Dyna class of architectures, named Introspective-Dyna, which is aimed at integrating reacting, planning and learning with self-awareness of an agent a belief subsystem able to help the agent having an attitude which emerges from within its internal representation of the world. An introspective prediction process is aimed at replacing a belief inferential process. Appropriateness (as soundness) and usefulness (as completeness) of taking an action are judged by means of faithfulness and fulfilment factors, as introspective explanative components of current decision making step. A double-sided temporal difference method is underlying each prediction step: the agent must look-ahead one step and choose an action with most-reachable anticipated rewards (real prediction) and must look backward (at least) one step and choose the action with most-acceptable introspectively anticipated results (introspective prediction). Two version of the I-Dyna algorithm, synchronous and asynchronous, are presented and discussed.
Palavras-chave: Reinforcement learning; adaptive reasoning; adaptive neural network; autonomous agents
Código DOI: 10.21528/CBRN1994-paper59
Artigo em PDF: CBRN1994-paper59.pdf
Arquivo BibTex: CBRN1994-paper59.bib