A self organising map (SOM) falls under the unsupervised machine learning category. It takes as input the number of neurons in the feature space and the number of epochs, which represent the iterations until the input fits a particular attribute category. A learning rate will affect how much displacement between each epoch the neurons will have. The result is a cluster of output were the algorithm tries to identify clusters of data and categorise it with other similar neurons forming part of this cluster. This produces a low-dimensional (two-dimensional) approximate representation of the input space of the attributes fed, known as a map. The major difference between other artificial neural networks is the competitive learning technique as opposed to error-correction learning. On the other hand, a disadvantage of SOM is the time it takes to prepare a model, and it is also hard to train against slowly evolving data.
A research paper highlights the use of SOMs and their potential use in dynamic difficulty adjustment for strategy games. Using SOM, it would be possible to understand the player’s decision making techniques when confronted with different scenarios, and from this data the AI could learn to overcome its strategy weaknesses.