Lexicographic Multi-Objective Reinforcement Learning

Authors: Joar Skalse, Lewis Hammond, Charlie Griffin, Alessandro Abate

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we evaluate our algorithms empirically. We first show how the learning time of LRL scales with the number of reward functions. We then compare the performance of VB-LRL and PB-LRL against that of other algorithms for solving constrained RL problems. Further experimental details and additional experiments are described in the supplementary material, and documented in our codebase.
Researcher Affiliation Academia Department of Computer Science, University of Oxford {joar.skalse, lewis.hammond, charlie.griffin, aabate}@cs.ox.ac.uk
Pseudocode Yes Algorithm 1 Lexicographic ϵ-Greedy, Algorithm 2 Value-Based Lexicographic RL, Algorithm 3 Policy-Based Lexicographic RL
Open Source Code Yes Further experimental details and additional experiments are described in the supplementary material, and documented in our codebase.5 Available at https://github.com/lrhammond/lmorl.
Open Datasets Yes The Cart Safe environment from gym-safety6 is a version of the classic Cart Pole environment.6 Available at https://github.com/jemaw/gym-safety. The Grid Nav environment, again from gym-safety (based on an environment in [Chow et al., 2018]), is a large gridworld... Finally, in the Intersection environment from highway-env7 the agent must guide a car through an intersection with dense traffic.7 Available at https://github.com/eleurent/highway-env
Dataset Splits No The paper mentions training models and evaluating performance but does not specify any explicit train/validation/test dataset splits or their percentages/counts.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU/CPU models, processor types, or memory amounts.
Software Dependencies No The paper mentions software environments like 'gym-safety' and 'highway-env' but does not specify any version numbers for these or other software dependencies.
Experiment Setup No The paper provides some high-level experimental details such as environment names and the number of runs, but it does not include concrete hyperparameter values, specific training configurations, or system-level settings within the main text. It defers further details to supplementary material.