Player Movement Models for Video Game Level Generation

Authors: Sam Snodgrass, Santiago Ontañón

IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our approach by generating levels for Super Mario Bros. We compare our results against the original levels, a previous constrained sampling approach, and a previous approach that learned a combined player and level model. 5 Experimental Evaluation We test our approach by sampling maps for the classic video game, Super Mario Bros. The remainder of this section describes the chosen domain, elaborates on the experimental set-up, and reports our obtained results. Table 1 shows the results of our experiments.
Researcher Affiliation Academia Sam Snodgrass, Santiago Onta n on Department of Computer Science, Drexel University Philadelphia, PA USA sps74@drexel.edu, santi@cs.drexel.edu
Pseudocode Yes Algorithm 1 Violation Location Resampling(w, h, C)
Open Source Code Yes The code used to train our models and perform our experiments are available online1. 1bitbucket.org/Sam_Snodgrass/ijcai_2017
Open Datasets No We extracted play traces for 4 maps using the method outlined in Section 4, for a total of 2,685 frames. We extracted the play traces from a gameplay video posted online2 of a single human player playing through the game. 2youtube.com/watch?v=bNNwNPUzCMo. The paper does not provide concrete access to the extracted play traces dataset.
Dataset Splits No We test our approach by training an Md MC on the 4 maps for which we had play traces. We then sampled 100 maps using the VLR algorithm paired with only the playability constraint and 100 maps with the VLR algorithm paired with both the playability constraint and the likelihood constraint. The paper does not explicitly provide training/validation/test dataset splits for their core data (the 4 Super Mario Bros maps and their extracted traces).
Hardware Specification No The paper does not explicitly mention any hardware specifications (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions implementing methods like k-medoids clustering and using an A* agent, but does not provide specific software names with version numbers (e.g., Python 3.8, PyTorch 1.9) for reproducibility.
Experiment Setup Yes For our experiments, we set the parameters as follows: rowsplits = 14, the height of the maps; lookahead = 3; and using the network structure n3, seen in Figure 1, and falling back to n2, n1, and n0 as needed. For our experiments, we use the Actions and Surroundings player model to evaluate the likelihood of the agent’s path during sampling. We chose the minimum value of 0.15 based on preliminary experimental results. We performed k-medoids clustering (with k = 20) using the 5x5 windows surrounding the player in each frame as the objects to cluster.