Safe Neurosymbolic Learning with Differentiable Symbolic Execution
Authors: Chenxi Yang, Swarat Chaudhuri
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the method on a mix of synthetic tasks and real-world benchmarks. Our experiments show that DSE significantly outperforms the state-of-the-art DIFFAI method on these tasks. |
| Researcher Affiliation | Academia | Chenxi Yang, Swarat Chaudhuri The University of Texas at Austin |
| Pseudocode | Yes | Algorithm 1: Learning Safe, Optimal Parameter Mixtures (Agarwal et al., 2018; Le et al., 2019) |
| Open Source Code | Yes | 1Our implementation of DSE is available at https://github.com/cxyang1997/DSE. |
| Open Datasets | No | The paper describes generating training data by executing ground-truth programs on uniformly sampled inputs or from path planners, but does not provide concrete access information (link, DOI, citation for an existing public dataset) for a publicly available dataset. |
| Dataset Splits | No | The paper mentions training, test data loss, and evaluates safety using an abstract interpreter on initial states, but does not specify explicit training/validation/test splits by percentage or absolute counts. It states "test data loss by running it on 10000 initial states that were not seen during training." and "splits the initial condition into a certain number of boxes (204 for Cartpole, 10000 for the other benchmarks)", which implicitly indicates data usage but not formal splits for general reproducibility. |
| Hardware Specification | Yes | We ran all the experiments using a single-thread implementation on a Linux system with Intel Xeon Gold 5218 2.30GHz CPUs and Ge Force RTX 2080 Ti GPUs. (Please refer to Appendix A.7 for more training details.) |
| Software Dependencies | No | Our framework is built on top of Py Torch (Paszke et al., 2019). We use the Adam Optimizer (Kingma & Ba, 2014) for all the experiments with default parameters and a weight decay of 0.000001. We ran all the experiments using a single-thread implementation on a Linux system with Intel Xeon Gold 5218 2.30GHz CPUs and Ge Force RTX 2080 Ti GPUs. |
| Experiment Setup | Yes | We use the Adam Optimizer (Kingma & Ba, 2014) for all the experiments with default parameters and a weight decay of 0.000001. We set the maximum epoch for Thermostat, AC and Racetrack as 1500, 1200 and 6000. We set the early stop if the final loss has less than a 1% decrease over 200 epochs. |