ARCH-COMP20 Category Report: Artificial Intelligence and Neural Network Control Systems (AINNCS) for Continuous and Hybrid Systems Plants

Abstract

This report presents the results of a friendly competition for formal verification of continuous and hybrid systems with artificial intelligence (AI) components. Specifically, machine learning (ML) components in cyber-physical systems (CPS), such as feedforward neural networks used as feedback controllers in closed-loop systems are considered, which is a class of systems classically known as intelligent control systems, or in more modern and specific terms, neural network control systems (NNCS). We more broadly refer to this category as AI and NNCS (AINNCS). The friendly competition took place as part of the workshop Applied Verification for Continuous and Hybrid Systems (ARCH) in 2020. In the second edition of this AINNCS category at ARCH-COMP, four tools have been applied to solve seven different benchmark problems, (in alphabetical order): NNV, OVERT, ReachNN*, and VenMAS. This report is a snapshot of the current landscape of tools and the types of benchmarks for which these tools are suited. Due to the diversity of problems, lack of a shared hardware platform, and the early stage of the competition, we are not ranking tools in terms of performance, yet the presented results probably provide the most complete assessment of current tools for safety verification of NNCS.

OVERT: An Algorithm for Safety Verification of Neural Network Control Policies for Nonlinear Systems

“…For the purpose of demonstration, we have chosen four classical control systems as described below. These four examples were part of the ARCH-COMP AINNC 2020 competition (Johnson et al, 2020). We have also trained additional control policies to facilitate comparison.…”

“…The position, velocity, and acceleration of the ego vehicle are x 4 , x 5 , and x 6 . The continuous-time dynamics of the system (Johnson et al, 2020) are:…”

“…Similar to Johnson et al (2020), we choose a = −2, which models the lead vehicle decelerating, and a friction parameter µ = 10 −4 . The discretization timestep is set to ∆τ = 0.1.…”

When Cyber-Physical Systems Meet AI: A Benchmark, an Evaluation, and a Way Forward

“…There is a variety of policy updating methods, such as Deep deterministic policy gradient (DDPG) [17], Twin-delayed deep deterministic policy gradient (TD3) [10], Actor-critic (A2C) [6], Proximal policy optimization (PPO) [24], Soft actor-critic (SAC) [11] and etc. We refer the readers to [16] for more details. We could see that DRL can be naturally applied to learn control policies for CPS by interacting with its plant, which is the physical environment of the system.…”

“…An annual workshop, namely ARCH, aims to mitigate this problem by bringing together CPS benchmarks and holding competitions 2 for different research topics. The most relevant competitions to this paper are Artificial Intelligence and Neural Network Control Systems [16] and Falsification [9]. However, the benchmark in the second competition only includes traditional CPS rather than AI-enabled CPS.…”