Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Evaluating Strategic Structures in Multi-Agent Inverse Reinforcement Learning

Authors: Justin Fu, Andrea Tacchetti, Julien Perolat, Yoram Bachrach

JAIR 2021 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our machinery on a classic game theory domain, a physics-based adversarial game, and a larger-scale simulated auction experiment where we show our method can extract accurate valuations for several popular auction mechanisms.
Researcher Affiliation Collaboration Justin Fu EMAIL University of California, Berkeley Department of Electrical Engineering & Computer Science Berkeley, CA, 94720, USA Andrea Tacchetti EMAIL Julien Perolat EMAIL Yoram Bachrach EMAIL Deep Mind 6 Pancras Square London, N1C 4AG, United Kingdom
Pseudocode Yes Algorithm 1 Inverse Equilibrium Single-Agent Reduction (IESAR) Method Input: Demonstration samples ˆσ πE1:N (Markov Games) Estimate πE1:N from ˆσ. for player i = 1 to N do Solve the single agent IRL problem with a utility-matching IRL method ˆRi = IRL(πE i |πE i) end for return ˆR1:N
Open Source Code No The paper does not explicitly provide a link to source code, state that code is released, or mention code in supplementary materials.
Open Datasets No The paper uses
Dataset Splits No The paper describes how demonstrations were obtained by simulating games and sampling from policies, or drawing valuations from a Gaussian distribution, but it does not specify explicit training/test/validation dataset splits of pre-existing datasets. For instance:
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments. It mentions using
Software Dependencies No The paper mentions several software components like
Experiment Setup Yes For all methods, we use a learning rate of 10 3 (selected via grid search between 10 4 and 10 1) for both the policy and utility functions. We constrain the norm of the utility function to c = 100 (selected via grid search between 1 and 1000). In our coordinate descent procedure, we optimize the inner loop (policy optimization) for 10 steps for each outer loop (utility optimization) step. ... For the environment, we used a horizon length of 25 steps...