Design & Technical Approach

Neural Networks Computational model based on structure and function of a biological neural network. Classifies game states in an adversarial setting, i.e. workers killed, buildings built, buildings destroyed, etc.

Generative Adversarial Networks Recently researched (Goodfellow et.al. 2014) algorithm to diversify input feature sets - allowing more DL insight to the data rather than generic constraints. Backpropagation takes place over the gradient of the Expectation (E) of the outputs of each of the Discriminator(D) and the Generator (G) networks. The Generator begins with a noise vector achieved from a Gaussian Parzen window - and adjusts that input vector based on how the Discriminator reacts to it - which simply reports it's KL-Divergence.

Unsupervised Machine Learning Machine Learning paradigm that involves attempting to learn datasets without explicit input-output pairs - rather we find an expected value and attempt to maximize that as much as D would allow. Utilizes unlabeled data to evaluate and conclude with feature domains. Algorithm comprehended game state events, incorporated weights to each, and outputted with associated areas for improvement per user, given replay data.

SC2Reader A python library that extracts data from various Starcraft II resources to power tools and services for the SC2 community. These include Replays, Maps, and Game Summaries. Retrieved information is converted to python-readable JSON format.