Powered by OpenAIRE graph
Found an issue? Give us feedback
image/svg+xml art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos Open Access logo, converted into svg, designed by PLoS. This version with transparent background. http://commons.wikimedia.org/wiki/File:Open_Access_logo_PLoS_white.svg art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos http://www.plos.org/ Archivio istituziona...arrow_drop_down
image/svg+xml art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos Open Access logo, converted into svg, designed by PLoS. This version with transparent background. http://commons.wikimedia.org/wiki/File:Open_Access_logo_PLoS_white.svg art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos http://www.plos.org/
addClaim

Deep Learning Approaches to Goal Recognition

Authors: Chiari, Mattia;

Deep Learning Approaches to Goal Recognition

Abstract

Riconoscere il goal di un agente utilizzando una traccia di osservazioni è un compito importante con diverse applicazioni. In letteratura, molti approcci di goal recognition (GR) si basano sull'applicazione di tecniche di pianificazione automatica che richiedono un modello delle azioni del dominio e dello stato iniziale del dominio (scritto, ad esempio, in PDDL). In questa tesi studiamo tre approcci alternativi (GRNet, Fast and Slow Goal Recognition e un approccio basato su BERT) in cui il goal recognition è formulato come un compito di classificazione affrontato utilizzando il machine learning. Tutti questi approcci mirano principalmente a risolvere istanze di GR in un dato dominio, specificato da un insieme di proposizioni e da un insieme di nomi di azioni. In GRNet, le istanze di classificazione del dominio sono risolte da una rete LSTM. L'unica informazione richiesta come input della rete addestrata è una traccia di nomi di azioni, ognuno dei quali indica solo il nome di un'azione osservata. Un'esecuzione della LSTM elabora una traccia di azioni osservate per calcolare la probabilità che ogni proposizione del dominio faccia parte del goal dell'agente. Fast and Slow Goal Recognition, ispirato al framework ``Thinking Fast and Slow'', è un modello a doppio processo che integra l'uso delle sopra-citate reti LSTM con le tecniche di pianificazione automatica. Questa architettura può sfruttare sia il riconoscimento veloce dei goal, basato sull'esperienza, fornito dalla rete, sia l'analisi lenta e deliberata fornita dalle tecniche di pianificazione. Infine, studiamo come un modello BERT addestrato sui piani sia in grado di comprendere il funzionamento di un dominio, le sue azioni e le loro relazioni reciproche. Questo modello viene poi sottoposto a fine-tuning per classificare le istanze di goal recognition. Le analisi sperimentali confermano che le architetture presentate raggiungono buone prestazioni sia in termini di accuratezza della classificazione dei goal che di tempo di esecuzione, ottenendo spesso risultati migliori rispetto a un sistema di goal recognition allo stato dell'arte sui benchmark considerati.

Recognising the goal of an agent from a trace of observations is an important task with many applications. In the literature, many approaches to goal recognition (GR) rely on the application of automated planning techniques which requires a model of the domain actions and of the initial domain state (written, e.g., in PDDL). We study three alternative approaches (GRNet, Fast and Slow Goal Recognition and a BERT-based approach) where Goal Recognition is formulated as a classification task addressed by machine learning. All these approaches are primarily aimed at solving GR instances in a given domain, which is specified by a set of propositions and a set of action names. In GRNet, the goal classification instances in the domain are solved by an LSTM network. The only information required as input of the trained network is a trace of action names, each one indicating just the name of an observed action. A run of the LSTM processes a trace of observed actions to compute how likely it is that each domain proposition is part of the agent's goal. Fast and Slow Goal Recognition, inspired by the ``Thinking Fast and Slow'' framework, is a dual-process model which integrates the use of the aforementioned LSTM with the automated planning techniques. This architecture can exploit both the fast, experience-based goal recognition provided by the network, and slow, deliberate analysis provided by the planning techniques. Finally, we study how a BERT model trained on plans is able to understand how a domain works, its actions and how they are related to each other. This model is then fine-tuned in order to classify goal recognition instances. Experimental analyses confirms that the presented architectures achieve good performance in terms of both goal classification accuracy and runtime, often obtaining better results w.r.t. a state-of-the-art GR system over the considered benchmarks.

Country
Italy
Related Organizations
  • BIP!
    Impact byBIP!
    selected citations
    These citations are derived from selected sources.
    This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    0
    popularity
    This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
    Average
    influence
    This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    Average
    impulse
    This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
    Average
Powered by OpenAIRE graph
Found an issue? Give us feedback
selected citations
These citations are derived from selected sources.
This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Citations provided by BIP!
popularity
This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
BIP!Popularity provided by BIP!
influence
This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Influence provided by BIP!
impulse
This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
BIP!Impulse provided by BIP!
0
Average
Average
Average
Green