Joint-Perturbation Simultaneous Pseudo-Gradient

Authors: Carlos Martin, Tuomas Sandholm

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on various games, including auctions, which have important real-world applications. Our approach yields a dramatic improvement in performance in terms of the wall time required to reach an approximate Nash equilibrium. We test our approach against the classical approach on several continuous-action games, in particular on many kinds of auction and on continuous Goofspiel (which can be thought of as a kind of sequential auction with budget constraints).
Researcher Affiliation Collaboration 1Carnegie Mellon University 2Strategy Robot, Inc. 3Optimized Markets, Inc. 4Strategic Machine, Inc.
Pseudocode No The paper describes the proposed method mathematically and in prose (Section 4 'Proposed Method'), including identities and estimators, but it does not present a structured pseudocode or algorithm block.
Open Source Code No The paper lists software used for the implementation ("For our code, we used Python 3.12.3, JAX 0.4.30 [Bradbury and others, 2018], Flax 0.8.5 [Heek and others, 2023], Optax 0.2.3 [Deep Mind and others, 2020], Matplotlib 3.9.1 [Hunter, 2007], and Sci Py 1.14.0 [Virtanen and others, 2020]") but does not explicitly state that the authors' own code for the methodology is open-source or provide a link to a code repository.
Open Datasets No For our experiment, we use a prior where bidder-item valuations vij are independently sampled uniformly at random from the unit interval. In our experiment, we sample the vis and cis from the standard uniform distribution, and sample C from the standard uniform distribution on [0, n]. The paper describes generating synthetic data for various auction scenarios by sampling from uniform distributions, rather than using a pre-existing public dataset with specific access information.
Dataset Splits No The paper describes generating data from specified distributions for its experiments, rather than using a fixed dataset that would then be split into training, validation, and test sets. Therefore, there is no explicit mention of dataset splits like percentages or sample counts for reproduction.
Hardware Specification Yes Each experiment was run individually on one NVIDIA A100 SXM4 40GB GPU.
Software Dependencies Yes For our code, we used Python 3.12.3, JAX 0.4.30 [Bradbury and others, 2018], Flax 0.8.5 [Heek and others, 2023], Optax 0.2.3 [Deep Mind and others, 2020], Matplotlib 3.9.1 [Hunter, 2007], and Sci Py 1.14.0 [Virtanen and others, 2020].
Experiment Setup Yes Our experimental hyperparameters are as follows. For each experiment, we run 8 trials. We use a stepsize of 10 4. For the Gaussian smoothing, we use a perturbation scale σ of 10 1. We use a batch size per iteration of 256. To update parameters, we use the Ada Belief optimizer [Zhuang and others, 2020]. For each player s strategy network, we use a single hidden layer of size 64, the Re LU activation function, and He initialization [He and others, 2015] for initializing the network s weights.