Subgame solving without common knowledge
Authors: Brian Zhang, Tuomas Sandholm
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted experiments on various small and mediumsized games to test the practical performance of 1-KLSS. |
| Researcher Affiliation | Collaboration | Brian Hu Zhang Computer Science Department Carnegie Mellon University bhzhang@cs.cmu.edu Tuomas Sandholm Computer Science Department, CMU Strategic Machine, Inc. Strategy Robot, Inc. Optimized Markets, Inc. sandholm@cs.cmu.edu |
| Pseudocode | Yes | Pseudocode for all algorithms can be found in the appendix. |
| Open Source Code | No | The paper mentions 'We have compiled and uploaded some representative samples of gameplay of our dark chess agent, with comments, at this link.' (Section 7) which refers to gameplay samples, not the source code for the methodology. No other explicit statement or link for the paper's own source code was found. |
| Open Datasets | No | The paper refers to specific games like 'Kuhn poker' and 'dark chess' and mentions 'some implemented in the open-source library Open Spiel [19]'. However, it does not provide concrete access information (e.g., specific links, DOIs, or citations with author and year) for the datasets or game instances used for the experiments. |
| Dataset Splits | No | The paper does not provide specific percentages, sample counts, or explicit methodology for training, validation, and test splits of the data used in its experiments. |
| Hardware Specification | Yes | Our agent runs on a single machine with 6 CPU cores. |
| Software Dependencies | Yes | All games were solved with Gurobi 9.0 [11] |
| Experiment Setup | Yes | To do this, we created a blueprint strategy for that is intentionally weak by forcing to play an ε-uniform strategy (i.e., at every infoset I, every action a must be played with probability at least ε/m where m is the number of actions at I). The blueprint is computed as the least-exploitable strategy under this condition. During subgame solving, the same restriction is applied at every infoset except the root, which means theoretically that it is possible for any strategy to arise from nested solving applied to every infoset in the game. The value ε was set to 0.25 in all experiments... |