Reinforcement Learning for Online Testing of Autonomous Driving Systems: a Replication and Extension Study

In a recent study, Reinforcement Learning (RL) used in combination with many-objective search, has been shown to outperform alternative techniques (random search and many-objective search) for online testing of Deep Neural Network-enabled systems. The empirical evaluation of these techniques was conducted on a state-of-the-art Autonomous Driving System (ADS). This work is a replication and extension of that empirical study. Our replication shows that RL does not outperform pure random test generation in a comparison conducted under the same settings of the original study, but with no confounding factor coming from the way collisions are measured. Our extension aims at eliminating some of the possible reasons for the poor performance of RL observed in our replication: (1) the presence of reward components providing contrasting or useless feedback to the RL agent; (2) the usage of an RL algorithm (Q-learning) which requires discretization of an intrinsically continuous state space. Results show that our new RL agent is able to converge to an effective policy that outperforms random testing. Results also highlight other possible improvements, which open to further investigations on how to best leverage RL for online ADS testing.

Boundary State Generation for Testing and Improvement of Autonomous Driving Systems

Recent advances in Deep Neural Networks (DNNs) and sensor technologies are enabling autonomous driving systems (ADSs) with an ever-increasing level of autonomy. However, assessing their dependability remains a critical concern. State-of-the-art ADS testing approaches modify the controllable attributes of a simulated driving environment until the ADS misbehaves. Such approaches have two main drawbacks: (1) modifications to the simulated environment might not be easily transferable to the in-field test setting (e.g., changing the road shape); (2) environment instances in which the ADS is successful are discarded, despite the possibility that they could contain hidden driving conditions in which the ADS may misbehave. In this paper, we present GenBo (GENerator of BOundary state pairs), a novel test generator for ADS testing. GenBo mutates the driving conditions of the ego vehicle (position, velocity and orientation), collected in a failure-free environment instance, and efficiently generates challenging driving conditions at the behavior boundary (i.e., where the model starts to misbehave) in the same environment. We use such boundary conditions to augment the initial training dataset and retrain the DNN model under test. Our evaluation results show that the retrained model has up to 16 higher success rate on a separate set of evaluation tracks with respect to the original DNN model.

Testing of Deep Reinforcement Learning Agents with Surrogate Models

Deep Reinforcement Learning (DRL) has received a lot of attention from the research community in recent years. As the technology moves away from game playing to practical contexts, such as autonomous vehicles and robotics, it is crucial to evaluate the quality of DRL agents. In this paper, we propose a search-based approach to test such agents. Our approach, implemented in a tool called Indago, trains a classifier on failure and non-failure environment configurations resulting from the DRL training process. The classifier is used at testing time as a surrogate model for the DRL agent execution in the environment, predicting the extent to which a given environment configuration induces a failure of the DRL agent under test. Indeed, the failure prediction acts as a fitness function, in order to guide the generation towards failure environment configurations, while saving computation time by deferring the execution of the DRL agent in the environment to those configurations that are more likely to expose failures. Experimental results show that our search-based approach finds 50% more failures of the DRL agent than state-of-the-art techniques. Moreover, such failure environment configurations, as well as the behaviours of the DRL agent induced by them, are significantly more diverse.

Two is Better Than One: Digital Siblings to Improve Autonomous Driving Testing

Simulation-based testing represents an important step to ensure the reliability of autonomous driving software. In practice, when companies rely on third-party general-purpose simulators, either for in-house or outsourced testing, the generalizability of testing results to real autonomous vehicles is at stake. In this paper, we strengthen simulation-based testing by introducing the notion of digital siblings, a novel framework in which the AV is tested on multiple general-purpose simulators, built with different technologies. First, test cases are automatically generated for each individual simulator. Then, tests are migrated between simulators, using feature maps to characterize of the exercised driving conditions. Finally, the joint predicted failure probability is computed and a failure is reported only in cases of agreement among the siblings. We implemented our framework using two open-source simulators and we empirically compared it against a digital twin of a physical scaled autonomous vehicle on a large set of test cases. Our study shows that the ensemble failure predictor by the digital siblings is superior to each individual simulator at predicting the failures of the digital twin. We discuss several ways in which our framework can help researchers interested in automated testing of autonomous driving software.

Adopting Two Supervisors for Efficient Use of Large-Scale Remote Deep Neural Networks

Recent decades have seen the rise of large-scale Deep Neural Networks (DNNs) to achieve human-competitive performance in a variety of artificial intelligence tasks. Often consisting of hundreds of millions, if not hundreds of billion parameters, these DNNs are too large to be deployed to, or efficiently run on resource-constrained devices such as mobile phones or IoT microcontrollers. Systems relying on large-scale DNNs thus have to call the corresponding model over the network, leading to substantial costs for hosting and running the large-scale remote model, costs which are often charged on a per-use basis. In this paper, we propose BiSupervised, a novel architecture, where, before relying on a large remote DNN, a system attempts to make a prediction on a small-scale local model. A DNN supervisor monitors said prediction process and identifies easy inputs for which the local prediction can be trusted. For these inputs, the remote model does not have to be invoked, thus saving costs, while only marginally impacting the overall system accuracy. Our architecture furthermore foresees a second supervisor to monitor the remote predictions and identify inputs for which not even these can be trusted, allowing to raise an exception or run a fallback strategy instead. We evaluate the cost savings, and the ability to detect incorrectly predicted inputs on four diverse case studies: IMDB movie review sentiment classification, Github issue triaging, Imagenet image classification, and SQuADv2 free-text question answering

When and Why Test Generators for Deep Learning Produce Invalid Inputs: an Empirical Study

Testing Deep Learning (DL) based systems inherently requires large and representative test sets to evaluate whether DL systems generalise beyond their training datasets. Diverse Test Input Generators (TIGs) have been proposed to produce artificial inputs that expose issues of the DL systems by triggering misbehaviours. Unfortunately, such generated inputs may be invalid, i.e., not recognisable as part of the input domain, thus providing an unreliable quality assessment. Automated validators can ease the burden of manually checking the validity of inputs for human testers, although input validity is a concept difficult to formalise and, thus, automate. In this paper, we investigate to what extent TIGs can generate valid inputs, according to both automated and human validators. We conduct a large empirical study, involving 2 different automated validators, 220 human assessors, 5 different TIGs and 3 classification tasks. Our results show that 84% artificially generated inputs are valid, according to automated validators, but their expected label is not always preserved. Automated validators reach a good consensus with humans (78% accuracy), but still have limitations when dealing with feature-rich datasets.

Uncertainty Quantification for Deep Neural Networks: An Empirical Comparison and Usage Guidelines

Deep Neural Networks (DNN) are increasingly used as components of larger software systems that need to process complex data, such as images, written texts, audio/video signals. DNN predictions cannot be assumed to be always correct for several reasons, among which the huge input space that is dealt with, the ambiguity of some inputs data, as well as the intrinsic properties of learning algorithms, which can provide only statistical warranties. Hence, developers have to cope with some residual error probability. An architectural pattern commonly adopted to manage failure-prone components is the supervisor, an additional component that can estimate the reliability of the predictions made by untrusted (e.g., DNN) components and can activate an automated healing procedure when these are likely to fail, ensuring that the Deep Learning based System (DLS) does not cause damages, despite its main functionality being suspended. In this paper, we consider DLS that implement a supervisor by means of uncertainty estimation. After overviewing the main approaches to uncertainty estimation and discussing their pros and cons, we motivate the need for a specific empirical assessment method that can deal with the experimental setting in which supervisors are used, where the accuracy of the DNN matters only as long as the supervisor lets the DLS continue to operate. Then we present a large empirical study conducted to compare the alternative approaches to uncertainty estimation. We distilled a set of guidelines for developers that are useful to incorporate a supervisor based on uncertainty monitoring into a DLS.

A Forgotten Danger in DNN Supervision Testing: Generating and Detecting True Ambiguity

Deep Neural Networks (DNNs) are becoming a crucial component of modern software systems, but they are prone to fail under conditions that are different from the ones observed during training (out-of-distribution inputs) or on inputs that are truly ambiguous, i.e., inputs that admit multiple classes with nonzero probability in their ground truth labels. Recent work proposed DNN supervisors to detect high-uncertainty inputs before their possible misclassification leads to any harm. To test and compare the capabilities of DNN supervisors, researchers proposed test generation techniques, to focus the testing effort on high-uncertainty inputs that should be recognized as anomalous by supervisors. However, existing test generators can only produce out-of-distribution inputs. No existing model- and supervisor-independent technique supports the generation of truly ambiguous test inputs. In this paper, we propose a novel way to generate ambiguous inputs to test DNN supervisors and used it to empirically compare several existing supervisor techniques. In particular, we propose AmbiGuess to generate ambiguous samples for image classification problems. AmbiGuess is based on gradient-guided sampling in the latent space of a regularized adversarial autoencoder. Moreover, we conducted what is - to the best of our knowledge - the most extensive comparative study of DNN supervisors, considering their capabilities to detect 4 distinct types of high-uncertainty inputs, including truly ambiguous ones.

Simple Techniques Work Surprisingly Well for Neural Network Test Prioritization and Active Learning (Replicability Study)

Test Input Prioritizers (TIP) for Deep Neural Networks (DNN) are an important technique to handle the typically very large test datasets efficiently, saving computation and labeling costs. This is particularly true for large-scale, deployed systems, where inputs observed in production are recorded to serve as potential test or training data for the next versions of the system. Feng et. al. propose DeepGini, a very fast and simple TIP, and show that it outperforms more elaborate techniques such as neuron- and surprise coverage. In a large-scale study (4 case studies, 8 test datasets, 32'200 trained models) we verify their findings. However, we also find that other comparable or even simpler baselines from the field of uncertainty quantification, such as the predicted softmax likelihood or the entropy of the predicted softmax likelihoods perform equally well as DeepGini.

Mind the Gap! A Study on the Transferability of Virtual vs Physical-world Testing of Autonomous Driving Systems

Safe deployment of self-driving cars (SDC) necessitates thorough simulated and in-field testing. Most testing techniques consider virtualized SDCs within a simulation environment, whereas less effort has been directed towards assessing whether such techniques transfer to and are effective with a physical real-world vehicle. In this paper, we leverage the Donkey Car open-source framework to empirically compare testing of SDCs when deployed on a physical small-scale vehicle vs its virtual simulated counterpart. In our empirical study, we investigate the transferability of behavior and failure exposure between virtual and real-world environments on a vast set of corrupted and adversarial settings. While a large number of testing results do transfer between virtual and physical environments, we also identified critical shortcomings that contribute to the reality gap between the virtual and physical world, threatening the potential of existing testing solutions when applied to physical SDCs.