Speculative Monte-Carlo Tree Search

Authors: Scott Cheng, Mahmut T Kandemir, Ding-Yong Hong

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical findings indicate that the proposed speculative MCTS can reduce training latency by 5.81 in 9x9 Go games. Moreover, our study shows that speculative execution can enhance the NN cache hit rate by 26% during midgame. Overall, our end-to-end evaluation indicates 1.91 speedup in 19x19 Go training time, compared to the state-of-the-art Kata Go program.
Researcher Affiliation Academia The Pennsylvania State University, USA Institute of Information Science, Academia Sinica, Taiwan
Pseudocode No The paper describes the speculative MCTS process with diagrams but does not include a formal pseudocode or algorithm block.
Open Source Code No The paper states: 'The model checkpoints are provided in https://doi.org/10.5281/ zenodo.13929490.' This provides model checkpoints, not the source code for the methodology itself.
Open Datasets No The paper mentions evaluating on '9x9 No Go [24], 9x9 Go, and 19x19 Go games' and refers to 'self-play simulation' and input features based on prior works [26, 27], but does not provide concrete access (link, DOI, or specific citation for public access) to the game data or records generated/used for training.
Dataset Splits No The paper mentions 'self-play simulation is 800' and 'measure the latency by averaging the results over 100,000 game moves', but it does not specify explicit training, validation, and test *data splits* (e.g., percentages or sample counts) for any dataset.
Hardware Specification Yes Our experiments are conducted on a cluster that consists of 8 NVIDIA V100 GPUs per node.
Software Dependencies No The paper mentions using 'Res Net blocks [25]' and 'SymPy [30]' for symbolic computing in Python, but it does not specify exact version numbers for any software libraries, frameworks, or programming languages used in their experimental setup.
Experiment Setup Yes Additionally, all models are trained from randomly initialized weights, and the self-play simulation is 800. ... In all our evaluations, the batch sizes for training and inference are set to 1024 and 24, respectively.