Scalable Online Exploration via Coverability

Authors: Philip Amortila, Dylan J Foster, Akshay Krishnamurthy

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we find that L1-Coverage effectively drives off-the-shelf policy optimization algorithms to explore the state space. We present proof-of-concept experiments to validate our theoretical results.
Researcher Affiliation Collaboration 1University of Illinois, Urbana-Champaign. 2Microsoft Research.
Pseudocode Yes Algorithm 1 Approximate Policy Cover Computation via L -Coverability Relaxation. Algorithm 2 Coverage-Driven Exploration (CODEX).
Open Source Code Yes Code available at github.com/philip-amortila/l1-coverability.
Open Datasets Yes We focus on the planning problem (Section 4), and consider the classical Mountain Car environment (Brockman et al., 2016).
Dataset Splits No The paper describes the environment setup and data generation process (e.g., 'deterministic starting state,' 'discretization'), but it does not specify explicit train/validation/test dataset splits or percentages for a pre-existing dataset.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions software like 'PyTorch (Paszke et al., 2019)', 'Adam optimizer (Kingma & Ba, 2015)', and 'Open AI Gym (Brockman et al., 2016)', but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We take a discount factor of 0.99, and a variance smoothing parameter of σ = 0.05. We train REINFORCE with horizons of length 400. We take πt, the policy which approximates Line 4 of Algorithm 1, to be the policy returned after 1000 REINFORCE updates, with one update after each rollout. The update in REINFORCE use the Adam optimizer (Kingma & Ba, 2015) with a learning rate of 10 3. We estimate all occupancies with N = 100 rollouts of length H = 200. We train for 20 epochs, corresponding to T = 20 in the loop of Line 3 of Algorithm 1. For the regularized reward of Eq. (16), we take ε = 10 4.