Programmatic Strategies for Real-Time Strategy Games

Authors: Julian R. H. Mariño, Rubens O. Moraes, Tassiana C. Oliveira, Claudio Toledo, Levi H. S. Lelis381-389

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate LS2 on four maps of µRTS, a real-time strategy (RTS) game designed for research purposes. We enlisted four professional programmers to write scripts for all four maps used in our experiment. In addition to the scripts written by programmers, we also compare the LS2 s scripts with search algorithms and other scripts from the µRTS codebase in a tournament-style experiment. A script LS2 synthesized obtained the highest winning rate in two of the maps tested and a script written by one of the programmers obtained the highest winning rate in the other two maps. In addition to this quantitative analysis, we also performed a qualitative analysis of the interpretability of the LS2 scripts, showing that the scripts synthesized in our experiments can be interpretable.
Researcher Affiliation Academia Julian R. H. Mari no,1 Rubens O. Moraes,2 Tassiana C. Oliveira,2 Claudio Toledo,1 Levi H. S. Lelis3 1 Departamento de Sistemas de Computac ao, ICMC, Universidade de S ao Paulo, Brazil 2 Departamento de Inform atica, Universidade Federal de Vic osa, Brazil 3 Department of Computing Science, Alberta Machine Intelligence Institute (Amii), University of Alberta, Canada
Pseudocode Yes Algorithm 1 Self play with local search
Open Source Code Yes Our implementation of LS2 is available online.1 1https://github.com/julianmarino/LS2
Open Datasets Yes We evaluate LS2 on four maps of µRTS (Onta n on 2017) by comparing them with tree search algorithms and scripts written by programmers. ... We chose to use µRTS in our experiments because it has an active research community with competitions being organized (Onta n on et al. 2018), with all competing algorithms available in a single codebase.2 2https://github.com/santiontanon/microrts
Dataset Splits No The paper does not provide specific training/validation/test dataset splits for the data used in the experiments. The evaluation is based on tournament-style matches rather than traditional dataset splits.
Hardware Specification No The paper mentions "We use one machine with 56 cores for our experiments" and refers to "computational resources of the Center for Mathematical Sciences Applied to Industry (Ce MEAI)", "the cluster Jupiter from Universidade Federal de Vic osa", and "the clusters from the Compute Canada". These descriptions lack specific hardware models (e.g., CPU, GPU models, memory details).
Software Dependencies No The paper mentions µRTS as the problem domain and µLanguage as their DSL, but it does not specify version numbers for any ancillary software dependencies or libraries used in their implementation.
Experiment Setup Yes We use n = 400 for Algorithm 1, and m = 70 and i = 20 for our local search algorithm. ... Instead of 100 ms, we allow A3N 500 ms of planning time in the self play match that generates T.