Game Engine Learning from Video

Authors: Matthew Guzdial, Boyang Li, Mark O. Riedl

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we present results from two distinct evaluations meant to demonstrate the utility of our system. [...] For the first evaluation we compare the frames predicted by our learned engine against a naive baseline and a Convolutional Neural Net (CNN)... For the second evaluation we evaluate the application of the learned engine s knowledge to training a game playing agent... We present the results of this evaluation in Figure 3. [...] we find that our learned engine predicts frames significantly more similar to the true frame (Wilxocon paired test, p < 2.2e 16).
Researcher Affiliation Academia Matthew Guzdial, Boyang Li, Mark O. Riedl School of Interactive Computing, Georgia Institute of Technology mguzdial3@gatech.edu, boyangli@gatech.edu, riedl@cc.gatech.edu
Pseudocode Yes Algorithm 1: frame scan; Algorithm 2: Engine Search
Open Source Code No No, the paper does not explicitly state that its own methodology's source code is available or provide a link to a repository. It mentions using 'Open CV [Pulli et al., 2012], an open-source machine vision toolkit', but this refers to a third-party tool, not the authors' own implementation.
Open Datasets Yes We made use of gameplay footage that had been evaluated in prior work by Summerville et al. [2016]. In particular we drew on the two most different players and their respective gameplay videos which Summerville et al. describe as the speedrunner and explorer.
Dataset Splits No No, the paper mentions a 'common train and test procedure' and that the CNN was 'trained on the same training data' and that 'performance on the test set converged', but it does not explicitly define a separate validation dataset or its split.
Hardware Specification Yes While this is an offline process, meant to only run once, the running time can still be prohibitive with lower values of θ (running for up to two weeks for a single game level on a 2013 i Mac).
Software Dependencies No No, the paper mentions using 'Open CV [Pulli et al., 2012]' but does not provide a specific version number for this or any other software dependency.
Experiment Setup Yes In addition we set θ to zero, to ensure the best possible output engine. [...] We made use of the same convolutional neural network architecture described in the Ranzato et al. paper, a network with four convolutional layers with 128 filters each, each making use of relu activation. [...] Thus we went from 416x364 pixel RGB images to 104x91 pixel grayscale images...