Artificial Intelligence (AI) planning technology represents one of the most efficient solutions to guide goal driven agents in domain problems which are mostly characterized by problem solving features. The current proliferation of physical and virtual autonomous agents have therefore triggered a growing interest in the automatic acquisition of action models, which can take advantage of the observation capabilities of agents, which sense world states through their sensors. In this work we present a Neural Network (NN) based approach for learning AI planning action models by observation in noisy environments. The system learns by observing a set of execution patterns of the same action in different contexts. The perceptions of both pre/post action execution states are used to train the NN learning component and are assumed to be affected by noise, which could be due to inaccurate reading or malfunctioning of the sensors. Preliminary experiments shows that the proposed NN learning module seems to be more resilient to unforeseen situations with respect to traditional propositional-based approaches to action model learning.

Neural Network Based Approach for Learning Planning Action Models

Milani A.
Membro del Collaboration Group
;
Biondi G.
Membro del Collaboration Group
2019

Abstract

Artificial Intelligence (AI) planning technology represents one of the most efficient solutions to guide goal driven agents in domain problems which are mostly characterized by problem solving features. The current proliferation of physical and virtual autonomous agents have therefore triggered a growing interest in the automatic acquisition of action models, which can take advantage of the observation capabilities of agents, which sense world states through their sensors. In this work we present a Neural Network (NN) based approach for learning AI planning action models by observation in noisy environments. The system learns by observing a set of execution patterns of the same action in different contexts. The perceptions of both pre/post action execution states are used to train the NN learning component and are assumed to be affected by noise, which could be due to inaccurate reading or malfunctioning of the sensors. Preliminary experiments shows that the proposed NN learning module seems to be more resilient to unforeseen situations with respect to traditional propositional-based approaches to action model learning.
2019
978-3-030-24310-4
978-3-030-24311-1
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11391/1476497
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 2
  • ???jsp.display-item.citation.isi??? ND
social impact