Feint Behaviors and Strategies: Formalization, Implementation and Evaluation

Authors: Junyu Liu, Xiangjun Peng

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results show that our design of Feint behaviors can (1) greatly improve the game reward gains; (2) significantly improve the diversity of Multi-Player Games; and (3) only incur negligible overheads in terms of time consumption.
Researcher Affiliation Academia Junyu Liu Brown University liu_junyu@brown.edu Xiangjun Peng The Chinese University of Hong Kong xjpeng@cse.cuhk.edu.hk
Pseudocode Yes Algorithm 1 in Appendix E illustrates the pseudo-code for pre-computing available Feint behavior templates given a set of available attack behaviors B. Algorithm 2 in Appendix E shows the pseudo-code for composing available Dual-Behavior models with backward searches.
Open Source Code No The NeurIPS checklist states 'No' for open access to data and code, justifying that the contribution is a formalization and implementation is based on existing frameworks, not releasing their own specific implementation code.
Open Datasets Yes Our main testbed game environment is a multi-player boxing game, which is based on Open AI s open-source environment Multi-Agent Particle Environment [23], but with heavy additional implementation to create a physically realistic scenario. We also modify and extend a strategic real-world game, Alpha Star [3], which is widely used as the experimental testbed in recent studies of Reinforcement Learning studies [28, 19].
Dataset Splits No The paper specifies training iterations but does not explicitly mention validation dataset splits or cross-validation setup.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running the experiments.
Software Dependencies No The paper mentions that its implementation is based on 'Johannesack/tf2multiagentrl [1]', but does not specify exact version numbers for programming languages, libraries, or other key software dependencies.
Experiment Setup Yes All experiments for the two-player scenario are trained for 75,000 game iterations and all experiments for the six-player scenario are trained for 150,000 game iterations.