Player Persona base on Long Short Term Memory (LSTM)

In user interaction, DKT feels a little too much. In most of time long short term memory is already handle time sequence action prediction.

For example, in rock-paper-scissor game. Usually this is totally random. But in people, it got patterns. A guy who uses rock for most time, he may still use rock for the next round.

Richard Bartle in his taxonomy of player types. He summarized that there are 4 types of players in MMORPGs. Killers, Explorers, Achievers and socializers.

So, in this demo, I provide a pirate game, the player can choose 4 options to communicate with the NPCs, Kill, Talk, WalkAround and AskForQuest. The top right shows a radar chat that the model assumes the next possible option the player will do.

Implementation

200 players and 30 rounds data is randomly generated to use as training data. So the model is not based on real data, DO NOT directly use the model for real circumstances.

Then, it is much easier than DKT, the input size is exactly how many the actions. While DKT needs to double it.

Then you can check the result.

Note

The model uses randomly generated data rather than real-world datasets. Consequently, the model’s predictions are not accurate and are intended solely to demonstrate the technology within a browser environment.