Are Equivariant Equilibrium Approximators Beneficial?
Authors: Zhijian Duan, Yunxuan Ma, Xiaotie Deng
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we theoretically characterize the benefits and limitations of equivariant equilibrium approximators. For the benefits, we show that they enjoy better generalizability than general ones and can achieve better approximations when the payoff distribution is permutation-invariant. For the limitations, we discuss their drawbacks in terms of equilibrium selection and social welfare. Together, our results help to understand the role of equivariance in equilibrium approximators. |
| Researcher Affiliation | Academia | 1Center on Frontiers of Computing Studies, School of Computer Science, Peking University, Beijing, China 2Center for Multi Agent Research, Institute for AI, Peking University, Beijing, China. Correspondence to: Xiaotie Deng <xiaotie@pku.edu.cn> |
| Pseudocode | Yes | Algorithm 1 Example: learning NE/CCE approximator via minibatch SGD |
| Open Source Code | No | The paper does not contain any explicit statements about providing open-source code or links to a code repository. |
| Open Datasets | No | Data is generated on-site for each training epoch. ... Specifically, the training and testing data are sampled from a fixed distribution, where the ε parameter of the game is sampled from a uniform distribution with a minimum of 0.005 and a maximum of 0.01 (i.e., U(0.005, 0.01)). |
| Dataset Splits | No | The paper specifies training and testing data sizes ("65536 data points used for training and 1000 data points for testing purposes" and "training on 4096 data points and testing on 1000 data points") but does not mention a separate validation set or its split details. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments, such as CPU or GPU models, or memory specifications. |
| Software Dependencies | No | The paper mentions the "Adam optimizer" and "Re LU" as activation functions, but does not provide specific version numbers for any software libraries, optimizers, or frameworks. |
| Experiment Setup | Yes | We employ a 5-layer fully connected neural network as the NE approximator, with 512 nodes in each hidden layer. Re LU serves as the activation function for each hidden layer, and batch normalization is applied before activation. During training, the Nash approximation loss function is utilized, and the model is optimized using the Adam optimizer. A batch size of 1024 is employed, with 65536 data points used for training and 1000 data points for testing purposes. Data is generated on-site for each training epoch. The initial learning rate is set to 5 × 10−4, with a decay ratio of γ = 0.3 at the 60th and 80th epochs. |