Partially Observable Monte Carlo Planning (POMCP) is a powerful online algorithm that can generate approximate policies for large Partially Observable Markov Decision Processes. The online nature of this method supports scalability by avoiding complete policy representation. However, the lack of an explicit representation of the policy hinders interpretability. In this thesis, we propose a methodology based on Maximum Satisfiability Modulo Theory (MAX-SMT) for analyzing POMCP policies by inspecting their traces, namely, sequences of belief-action pairs generated by the algorithm. The proposed method explores local properties of the policy to build a compact and informative summary of the policy behaviour. This representation exploits a high-level description encoded using logical formulas that domain experts can provide. The final formula can be used to identify unexpected decisions, namely, decisions that violate the expert indications. We show that this identification process can be used offline (to improve the explainability of the policy and to identify anomalous behaviours) or online (to shield the decisions of the POMCP algorithm). We also present an active methodology that can effectively query a POMCP policy to build more reliable descriptions quickly. We extensively evaluate our methodologies on two standard benchmarks for POMDPs, namely, emph{tiger} and emph{rocksample}, and on a problem related to velocity regulation in mobile robot navigation. Results show that our approach achieves good performance due to its capability to exploit experts' knowledge of the domains. Specifically, our approach can be used both to identify anomalous behaviours in faulty POMCPs and to improve the performance of the system by using the shielding mechanism. In the first case, we test the methodology against a state-of-the-art anomaly detection algorithm, while in the second, we compared the performance of shielded and unshielded POMCPs. We implemented our methodology in CC, and the code is open-source and available at href{https://github.com/GiuMaz/XPOMCP}{https://github.com/GiuMaz/XPOMCP}.

Rule-Based Policy Interpretation and Shielding for Partially Observable Monte Carlo Planning

Giulio Mazzi
;
Alberto Castellini
Supervision
;
Alessandro Farinelli
Supervision
2022

Abstract

Partially Observable Monte Carlo Planning (POMCP) is a powerful online algorithm that can generate approximate policies for large Partially Observable Markov Decision Processes. The online nature of this method supports scalability by avoiding complete policy representation. However, the lack of an explicit representation of the policy hinders interpretability. In this thesis, we propose a methodology based on Maximum Satisfiability Modulo Theory (MAX-SMT) for analyzing POMCP policies by inspecting their traces, namely, sequences of belief-action pairs generated by the algorithm. The proposed method explores local properties of the policy to build a compact and informative summary of the policy behaviour. This representation exploits a high-level description encoded using logical formulas that domain experts can provide. The final formula can be used to identify unexpected decisions, namely, decisions that violate the expert indications. We show that this identification process can be used offline (to improve the explainability of the policy and to identify anomalous behaviours) or online (to shield the decisions of the POMCP algorithm). We also present an active methodology that can effectively query a POMCP policy to build more reliable descriptions quickly. We extensively evaluate our methodologies on two standard benchmarks for POMDPs, namely, emph{tiger} and emph{rocksample}, and on a problem related to velocity regulation in mobile robot navigation. Results show that our approach achieves good performance due to its capability to exploit experts' knowledge of the domains. Specifically, our approach can be used both to identify anomalous behaviours in faulty POMCPs and to improve the performance of the system by using the shielding mechanism. In the first case, we test the methodology against a state-of-the-art anomaly detection algorithm, while in the second, we compared the performance of shielded and unshielded POMCPs. We implemented our methodology in CC, and the code is open-source and available at href{https://github.com/GiuMaz/XPOMCP}{https://github.com/GiuMaz/XPOMCP}.
SMT, planning, POMDP, POMCP
File in questo prodotto:
File Dimensione Formato  
Tesi_dottorato_Mazzi.pdf

accesso aperto

Descrizione: Tesi dottorato - Giulio Mazzi
Tipologia: Tesi di dottorato
Licenza: Creative commons
Dimensione 1.83 MB
Formato Adobe PDF
1.83 MB Adobe PDF Visualizza/Apri

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/11562/1067927
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact