In sequential decision making, neural networks (NNs) are nowadays commonly used to represent and learn the agent’s policy. This area of application has implied new software quality assessment challenges that traditional validation and verification practises are not able to handle. Subsequently, novel approaches have emerged to adapt those techniques to NN-based policies for sequential decision making. This survey paper aims at summarising these novel contributions and proposing future research directions. We conducted a literature review of recent research papers (from 2018 to beginning of 2023), whose topics cover aspects of the test or verification of NN-based policies. The selection has been enriched by a snowballing process from the previously selected papers, in order to relax the scope of the study and provide the reader with insight into similar verification challenges and their recent solutions. 18 papers have been finally selected. Our results show evidence of increasing interest for this subject. They highlight the diversity of both the exact problems considered and the techniques used to tackle them.

Metamorphic Testing is a software testing paradigm which aims at using necessary properties of a system-under-test, called metamorphic relations, to either check its expected outputs, or to generate new test cases. Metamorphic Testing has been successful to test programs for which a full oracle is not available or to test programs for which there are uncertainties on expected outputs such as learning systems. In this article, we propose Adaptive Metamorphic Testing as a generalization of a simple yet powerful reinforcement learning technique, namely contextual bandits, to select one of the multiple metamorphic relations available for a program. By using contextual bandits, Adaptive Metamorphic Testing learns which metamorphic relations are likely to transform a source test case, such that it has higher chance to discover faults. We present experimental results over two major case studies in machine learning, namely image classification and object detection, and identify weaknesses and robustness boundaries. Adaptive Metamorphic Testing efficiently identifies weaknesses of the tested systems in context of the source test case.

Solving Constraint Optimization Problems (COP) requires exploring a large search space. By providing objective boundaries, this space can be pruned. Finding close boundaries, that correctly under- or overestimate the optimum, is difficult without having a heuristic function for the COP. We present a method for learning to estimate boundaries from problem instances using machine learning. The trained model can then estimate boundaries for unseen instances and thereby support the constraint solver through an additional boundary constraint.

Constraint Programming (CP) is a powerful declarative programming paradigm combining inference and search in order to find solutions to various type of constraint systems. Dealing with highly disjunctive constraint systems is notoriously difficult in CP. Apart from trying to solve each disjunct independently from each other, there is little hope and effort to succeed in constructing intermediate results combining the knowledge originating from several disjuncts. In this paper, we propose If Then Else (ITE), a lightweight approach for implementing stratified constructive disjunction and negation on top of an existing CP solver, namely SICStus Prolog clp(FD). Although constructive disjunction is known for more than three decades, it does not have straightforward implementations in most CP solvers. ITE is a freely available library proposing stratified and constructive reasoning for various operators, including disjunction and negation, implication and conditional. Our preliminary experimental results show that ITE is competitive with existing approaches that handle disjunctive constraint systems.