Usage of Deep Neural Networks in Game Data Science

Advantages and disadvantages of FNNs

Feedforward neural networks can be used to learn highly complex nonlinear relationships between the input data and the output. However, such flexibility comes at a cost, as NNs can overfit very easily. As such, it’s important to properly evaluate NNs and compare their performance to other simpler models to ensure that the resulting model didn’t overfit.

Vanishing gradient problem

Another challenge faced when training NNs is called the vanishing gradient problem, which occurs when gradient-based optimization techniques are used with certain activation functions. Specifically, as we move backward through the hidden layers during backpropogation, the gradients tend to get smaller and smaller (vanishingly small), preventing the weight from changing its value. This means that the neurons in the earlier layers learn very slowly because of the very small updates in their weights as compared with the neurons in the later layers of the network. As a result, the training time increases, and in the worst case, this may completely stop the NN from further training in the layers near the input. This problem makes it challenging to learn and tune the parameters of the earlier layers in the networks, especially if the network has large numbers of layers.

DNNs

There are several examples within the games industry and research that use NNs and specifically DNNs, for different types of predictions and item recommendations.

Example

Bertens et al. (2018) show a great example where they were concerned with developing a model that can predict the next item that a player would purchase in a game. In modern games, the performance and strategy of players depend on the types of items they use. Most games allow players to buy a wide range of virtual items with real money. Due to this open market of items, sometimes, players may be overwhelmed by the variety and the capability of different items they are presented. Therefore, the industry has produced many techniques for item recommendation. Correctly predicting items of interest for different players would result in a good recommendation system. To predict items players would purchase next, Bertens et al. proposed two different approaches: Extremely Randomized Tree (ERT) and Deep Neural Network.