Quality Diversity through Human Feedback: Towards Open-Ended Diversity-Driven Optimization

Authors: Li Ding, Jenny Zhang, Jeff Clune, Lee Spector, Joel Lehman

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical studies show that QDHF significantly outperforms state-of-the-art methods in automatic diversity discovery and matches the efficacy of QD with manually crafted diversity metrics on standard benchmarks in robotics and reinforcement learning.
Researcher Affiliation Collaboration 1Manning College of Information & Computer Sciences, University of Massachusetts Amherst 2Department of Computer Science, University of British Columbia 3Vector Institute 4Canada CIFAR AI Chair 5Department of Computer Science, Amherst College 6Stochastic Labs *Part of the work was done while the author was affiliated with Stability AI.
Pseudocode Yes Algorithm 1 MAP-Elites Algorithm 1: Initialize a map of solutions, each cell representing a unique feature combination 2: while not converged do 3: Generate new solutions via mutation and crossover 4: for each solution do 5: Evaluate the solution for its performance and feature characteristics 6: Identify the corresponding cell in the map based on features 7: if solution is better than the current cell occupant then 8: Replace the cell s solution with the new solution 9: end if 10: end for 11: end while 12: Return the map of elite solutions
Open Source Code Yes Code and tutorials are available at https://liding.info/qdhf.
Open Datasets Yes We use human judgment data from the NIGHTS dataset, and the Dream Sim model to estimate image similarity, both from Fu et al. (2023).
Dataset Splits No The paper mentions evaluating on a 'validation set' but does not provide specific details such as percentages, sample counts, or the methodology for creating these splits.
Hardware Specification No The paper mentions using a 'high-performance computing cluster at the Massachusetts Green High Performance Computing Center (MGHPCC)' but does not provide specific hardware details like GPU/CPU models or memory amounts.
Software Dependencies Yes We use Stable Diffusion v2.1-base, which generates images at a resolution of 512x512. The feature extractor is a CLIP model with Vi T-B/16 backbone, which returns a 512-dim feature vector.
Experiment Setup Yes For all experiments, we run MAPElites for 1000 iterations, and for each iteration, we generate a batch of 100 solutions with Gaussian mutation (adding Gaussian noises sampled from N(0, 0.12)), and evaluate them. The archive has a shape of (50, 50), i.e., each of the 2 dimensions is discretized into 50 equal-sized bins.