Infinite Time Horizon Safety of Bayesian Neural Networks
Authors: Mathias Lechner, Đorđe Žikelić, Krishnendu Chatterjee, Thomas Henzinger
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform an experimental evaluation of our proposed method for learning positive invariant neural networks that prove infinite time horizon safety. Our evaluation consists of an ablation study where we disable different core components of Algorithm 1 and measure their effects on the obtained safety bounds and the algorithm s runtime. |
| Researcher Affiliation | Academia | Mathias Lechner IST Austria Klosterneuburg, Austria mlechner@ist.ac.at Ðor de Žikeli c IST Austria Klosterneuburg, Austria dzikelic@ist.ac.at Krishnendu Chatterjee IST Austria Klosterneuburg, Austria kchatterjee@ist.ac.at Thomas A. Henzinger IST Austria Klosterneuburg, Austria tah@ist.ac.at |
| Pseudocode | Yes | The pseudocode is given in Algorithm 1. |
| Open Source Code | Yes | Code is publicly available 2. https://github.com/mlech26l/bayesian_nn_safety |
| Open Datasets | Yes | We evaluate our approach on a series of reinforcement learning benchmarks, including non-Lyapunovian safety specifications. ... We conduct our evaluation on three benchmark environments that differ in terms of complexity and safety specifications. ... Our first benchmark represents an unstable linear dynamical system... Our second benchmark is the inverted pendulum task... our third benchmark is a collision avoidance task... |
| Dataset Splits | No | No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning is provided in the main text. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) are mentioned in the paper's main text for running experiments. |
| Software Dependencies | Yes | the verification steps of our algorithm can be reduced to MILP-solving using Gurobi [20]. |
| Experiment Setup | Yes | We train two BNN policies for each benchmark-ablation pair, one with Bayesian weights from the second layer on (with N(0, 0.1) prior) and one with Bayesian weights in all layers (with N(0, 0.05) prior). |