Safeguarded Progress in Reinforcement Learning: Safe Bayesian Exploration for Control Policy Synthesis

This paper addresses the problem of maintaining safety during training in Reinforcement Learning (RL), such that the safety constraint violations are bounded at any point during learning. In a variety of RL applications the safety of the agent is particularly important, e.g. autonomous platforms or robots that work in proximity of humans. As enforcing safety during training might severely limit the agent's exploration, we propose here a new architecture that handles the trade-off between efficient progress and safety during exploration. As the exploration progresses, we update via Bayesian inference Dirichlet-Categorical models of the transition probabilities of the Markov decision process that describes the environment dynamics. This paper proposes a way to approximate moments of belief about the risk associated to the action selection policy. We construct those approximations, and prove the convergence results. We propose a novel method for leveraging the expectation approximations to derive an approximate bound on the confidence that the risk is below a certain level. This approach can be easily interleaved with RL and we present experimental results to showcase the performance of the overall architecture.

You Only Explain Once

In this paper, we propose a new black-box explainability algorithm and tool, YO-ReX, for efficient explanation of the outputs of object detectors. The new algorithm computes explanations for all objects detected in the image simultaneously. Hence, compared to the baseline, the new algorithm reduces the number of queries by a factor of 10X for the case of ten detected objects. The speedup increases further with with the number of objects. Our experimental results demonstrate that YO-ReX can explain the outputs of YOLO with a negligible overhead over the running time of YOLO. We also demonstrate similar results for explaining SSD and Faster R-CNN. The speedup is achieved by avoiding backtracking by combining aggressive pruning with a causal analysis.

Multiple Different Explanations for Image Classifiers

Existing explanation tools for image classifiers usually give only one single explanation for an image. For many images, however, both humans and image classifiers accept more than one explanation for the image label. Thus, restricting the number of explanations to just one severely limits the insight into the behavior of the classifier. In this paper, we describe an algorithm and a tool, REX, for computing multiple explanations of the output of a black-box image classifier for a given image. Our algorithm uses a principled approach based on causal theory. We analyse its theoretical complexity and provide experimental results showing that REX finds multiple explanations on 7 times more images than the previous work on the ImageNet-mini benchmark.

LCRL: Certified Policy Synthesis via Logically-Constrained Reinforcement Learning

LCRL is a software tool that implements model-free Reinforcement Learning (RL) algorithms over unknown Markov Decision Processes (MDPs), synthesising policies that satisfy a given linear temporal specification with maximal probability. LCRL leverages partially deterministic finite-state machines known as Limit Deterministic Buchi Automata (LDBA) to express a given linear temporal specification. A reward function for the RL algorithm is shaped on-the-fly, based on the structure of the LDBA. Theoretical guarantees under proper assumptions ensure the convergence of the RL algorithm to an optimal policy that maximises the satisfaction probability. We present case studies to demonstrate the applicability, ease of use, scalability, and performance of LCRL. Owing to the LDBA-guided exploration and LCRL model-free architecture, we observe robust performance, which also scales well when compared to standard RL approaches (whenever applicable to LTL specifications). Full instructions on how to execute all the case studies in this paper are provided on a GitHub page that accompanies the LCRL distribution www.github.com/grockious/lcrl.

Exposing Previously Undetectable Faults in Deep Neural Networks

Existing methods for testing DNNs solve the oracle problem by constraining the raw features (e.g. image pixel values) to be within a small distance of a dataset example for which the desired DNN output is known. But this limits the kinds of faults these approaches are able to detect. In this paper, we introduce a novel DNN testing method that is able to find faults in DNNs that other methods cannot. The crux is that, by leveraging generative machine learning, we can generate fresh test inputs that vary in their high-level features (for images, these include object shape, location, texture, and colour). We demonstrate that our approach is capable of detecting deliberately injected faults as well as new faults in state-of-the-art DNNs, and that in both cases, existing methods are unable to find these faults.

Compositional Explanations for Image Classifiers

Existing algorithms for explaining the output of image classifiers perform poorly on inputs where the object of interest is partially occluded. We present a novel, black-box algorithm for computing explanations that uses a principled approach based on causal theory. We implement the method in the tool CET (Compositional Explanation Tool). Owing to the compositionality in its algorithm, CET computes explanations that are much more accurate than those generated by the existing explanation tools on images with occlusions and delivers a level of performance comparable to the state of the art when explaining images without occlusions.

Neural Termination Analysis

We introduce a novel approach to the automated termination analysis of computer programs: we train neural networks to act as ranking functions. Ranking functions map program states to values that are bounded from below and decrease as the program runs. The existence of a valid ranking function proves that the program terminates. While in the past ranking functions were usually constructed using static analysis, our method learns them from sampled executions. We train a neural network so that its output decreases along execution traces as a ranking function would; then, we use formal reasoning to verify whether it generalises to all possible executions. We present a custom loss function for learning lexicographic ranking functions and use satisfiability modulo theories for verification. Thanks to the ability of neural networks to generalise well, our method succeeds over a wide variety of programs. This includes programs that use data structures from standard libraries. We built a prototype analyser for Java bytecode and show the efficacy of our method over a standard dataset of benchmarks.

Shielding Atari Games with Bounded Prescience

Deep reinforcement learning (DRL) is applied in safety-critical domains such as robotics and autonomous driving. It achieves superhuman abilities in many tasks, however whether DRL agents can be shown to act safely is an open problem. Atari games are a simple yet challenging exemplar for evaluating the safety of DRL agents and feature a diverse portfolio of game mechanics. The safety of neural agents has been studied before using methods that either require a model of the system dynamics or an abstraction; unfortunately, these are unsuitable to Atari games because their low-level dynamics are complex and hidden inside their emulator. We present the first exact method for analysing and ensuring the safety of DRL agents for Atari games. Our method only requires access to the emulator. First, we give a set of 43 properties that characterise "safe behaviour" for 30 games. Second, we develop a method for exploring all traces induced by an agent and a game and consider a variety of sources of game non-determinism. We observe that the best available DRL agents reliably satisfy only very few properties; several critical properties are violated by all agents. Finally, we propose a countermeasure that combines a bounded explicit-state exploration with shielding. We demonstrate that our method improves the safety of all agents over multiple properties.

Ranking Policy Decisions

Policies trained via Reinforcement Learning (RL) are often needlessly complex, making them more difficult to analyse and interpret. In a run with $n$ time steps, a policy will decide $n$ times on an action to take, even when only a tiny subset of these decisions deliver value over selecting a simple default action. Given a pre-trained policy, we propose a black-box method based on statistical fault localisation that ranks the states of the environment according to the importance of decisions made in those states. We evaluate our ranking method by creating new, simpler policies by pruning decisions identified as unimportant, and measure the impact on performance. Our experimental results on a diverse set of standard benchmarks (gridworld, CartPole, Atari games) show that in some cases less than half of the decisions made contribute to the expected reward. We furthermore show that the decisions made in the most frequently visited states are not the most important for the expected reward.