Simplifying Constraint Inference with Inverse Reinforcement Learning

Authors: Adriana Hugessen, Harley Wiltzer, Glen Berseth

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct a series of experiments across several environments in order to answer the following questions: (1) How does IRL perform on constraint inference tasks compared to Lagrangian methods? and (2) How do the proposed modifications or regularizations over vanilla IRL improve performance on constraint inference tasks?
Researcher Affiliation Academia Adriana Hugessen Mila, Université de Montréal Harley Wiltzer Mila, Mc Gill University Glen Berseth Mila, Université de Montréal
Pseudocode Yes Algorithm 1 IRL for ICRL Separate Critics
Open Source Code Yes Our code is made available at https://github.com/ahugs/simple-icrl.
Open Datasets Yes For our experiments, we consider the virtual environments for benchmarking inverse constraint learning, introduced by Liu et al. [2023a] since these were specially designed to test the performance of constraint inference tasks and also provide a recent baseline for Lagrangian-based constraint inference methods, including expert data.
Dataset Splits No We compare all methods according to average performance in the last 50 testing episodes and report statistics (IQM, Median, Mean and Optimality Gap) with bootstrapped 95% confidence intervals computed across five seeds according to the method recommended in Agarwal et al. [2021].
Hardware Specification No Each run (consisting of five seeds) was trained on a node with a single GPU (varying GPU resources were used), 6 CPUs and 6GB of RAM per CPU.
Software Dependencies Yes All of our code is based on the Tianshou [Weng et al., 2022] and FSRL [Liu et al., 2023b] implementations of SAC and SAC-Lagrangian, respectively. Here we include all the hyperparameter configurations for our experiments. Any hyperparameters not listed here use the default hyperparameters in their respective libraries (Tianshou version 1.0.0 and FSRL version 0.1.0).
Experiment Setup Yes All of our code is based on the Tianshou [Weng et al., 2022] and FSRL [Liu et al., 2023b] implementations of SAC and SAC-Lagrangian, respectively. Here we include all the hyperparameter configurations for our experiments. Any hyperparameters not listed here use the default hyperparameters in their respective libraries (Tianshou version 1.0.0 and FSRL version 0.1.0).