Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning

Authors: Marvin Alles, Philip Becker-Ehmck, Patrick van der Smagt, Maximilian Karl

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate C-LAP on the D4RL and V-D4RL benchmark, and show that C-LAP is competitive to state-of-the-art methods, especially outperforming on datasets with visual observations.
Researcher Affiliation Collaboration 1Machine Learning Research Lab, Volkswagen Group 2Technical University of Munich 3Eötvös Loránd University Budapest
Pseudocode Yes B Algorithm We provide the general algorithm of C-LAP. Algorithm 1: C-LAP
Open Source Code Yes 1Code is available at: https://github.com/marvinalles/c-lap
Open Datasets Yes We empirically evaluate C-LAP on the D4RL and V-D4RL benchmark, and show that C-LAP is competitive to state-of-the-art methods, especially outperforming on datasets with visual observations. ... To focus on the latter, we separately evaluate the performance on low-dimensional feature observations using the D4RL benchmark [26], and on image observations using the V-D4RL benchmark [20].
Dataset Splits No The paper describes using D4RL and V-D4RL benchmarks, which typically have predefined train/test splits. However, it does not explicitly state the dataset splits for training, validation, or testing within the paper's text, nor does it refer to predefined splits with specific citations or percentages. It only mentions 'static dataset D = {(o1:T , a1:T , r1:T )N n=1}' for offline reinforcement learning in general.
Hardware Specification Yes C-LAP experiments with visual observations take around 10 hours on a RTX8000 GPU and experiments with low-dimension feature observations around 11 hours on a A100 GPU.
Software Dependencies Yes We implement all methods in JAX [33] using Equinox [34].
Experiment Setup Yes We provide the hyper-parameters of CLAP in Table 2 and the constraint values used for the D4RL benchmark in Table 3 and for the V-D4RL benchmark in Table 4.