TY - GEN
T1 - Differential Safety Testing of Deep RL Agents Enabled by Automata Learning
AU - Tappler, Martin
AU - Aichernig, Bernhard K.
N1 - Publisher Copyright:
© 2024, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2024
Y1 - 2024
N2 - Learning-enabled controllers (LECs) pose severe challenges to verification. Their decisions often come from deep neural networks that are hard to interpret and verify, and they operate in stochastic and unknown environments with high-dimensional state space. These complexities make analyses of the internals of LECs and manual modeling of the environments extremely challenging. Numerous combinations of automata learning with verification techniques have shown its potential in the analysis of black-box reactive systems. Hence, automata learning may also bring light into the black boxes that are LECs and their runtime environments. A hurdle to the adoption of automata-learning-based verification is that it is often difficult to provide guarantees on the accuracy of learned automata. This is exacerbated in complex, stochastic environments faced by LECs. In this paper, we demonstrate that accuracy guarantees on learned models are not strictly necessary. Through a combination of automata learning, testing, and statistics, we perform testing-based verification with statistical guarantees in the absence of guarantees on the learned automata. We showcase our approach by testing deep reinforcement learning for safety that have been trained to play the computer game Super Mario Bros.
AB - Learning-enabled controllers (LECs) pose severe challenges to verification. Their decisions often come from deep neural networks that are hard to interpret and verify, and they operate in stochastic and unknown environments with high-dimensional state space. These complexities make analyses of the internals of LECs and manual modeling of the environments extremely challenging. Numerous combinations of automata learning with verification techniques have shown its potential in the analysis of black-box reactive systems. Hence, automata learning may also bring light into the black boxes that are LECs and their runtime environments. A hurdle to the adoption of automata-learning-based verification is that it is often difficult to provide guarantees on the accuracy of learned automata. This is exacerbated in complex, stochastic environments faced by LECs. In this paper, we demonstrate that accuracy guarantees on learned models are not strictly necessary. Through a combination of automata learning, testing, and statistics, we perform testing-based verification with statistical guarantees in the absence of guarantees on the learned automata. We showcase our approach by testing deep reinforcement learning for safety that have been trained to play the computer game Super Mario Bros.
KW - Automata Learning
KW - Differential Testing
KW - Learning-Based Testing
KW - Reinforcement Learning
UR - http://www.scopus.com/inward/record.url?scp=85180633679&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-46002-9_8
DO - 10.1007/978-3-031-46002-9_8
M3 - Conference paper
AN - SCOPUS:85180633679
SN - 9783031460012
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 138
EP - 159
BT - Bridging the Gap Between AI and Reality - 1st International Conference, AISoLA 2023, Proceedings
A2 - Steffen, Bernhard
PB - Springer Science and Business Media Deutschland GmbH
T2 - 1st International Conference on Bridging the Gap between AI and Reality
Y2 - 23 October 2023 through 28 October 2023
ER -