DAG-Based Column Generation for Adversarial Team Games

Authors: Youzhi Zhang, Bo An, Daniel Dajun Zeng

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experimental Evaluation We evaluate the performance of DCG and run all experiments on a machine with a 4-core 2.3GHz CPU (8 threads) and 16GB of RAM available by using CPLEX 20.1.
Researcher Affiliation Collaboration 1Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences 2Nanyang Technological University, Singapore 3Skywork AI, Singapore 4Institute of Automation, Chinese Academy of Sciences, Beijing, China.
Pseudocode Yes Algorithm 1 Expanding the TB-DAG
Open Source Code No No explicit statement or link providing access to the authors' source code was found. The paper mentions that 'the codes for these baselines are not available'.
Open Datasets Yes We then use two standard extensive-form games (Farina et al., 2018; 2021; Carminati et al., 2022): Kuhn poker and Leduc poker (details on them can be found in these references).
Dataset Splits No The paper discusses extensive-form games (Kuhn poker and Leduc poker) which are game environments, not typical datasets that are split into train/validation/test sets for machine learning models. No explicit mention of data splits was found.
Hardware Specification Yes We evaluate the performance of DCG and run all experiments on a machine with a 4-core 2.3GHz CPU (8 threads) and 16GB of RAM available by using CPLEX 20.1.
Software Dependencies Yes We evaluate the performance of DCG and run all experiments on a machine with a 4-core 2.3GHz CPU (8 threads) and 16GB of RAM available by using CPLEX 20.1.
Experiment Setup Yes For these CG-based algorithms, we randomly initialize the restricted game with a coordinated strategy for the team, and for the CG-based algorithms with semi-randomized strategies, we initialize a uniform strategy for the corresponding player. For each of these algorithms, we consider one more variant: at each iteration, it solves the linear relaxation of the BRO first to see if it can output the optimal solution added to the restricted game; if it cannot do that, it solves the original mixed-integer BRO and adds all feasible solutions for the BRO from the CPLEX solution pool to the original game. ...with target precision of the team value in a TMECor is 10^-6