The naturalAI project aims to develop AI so that it seems natural to interact with it. Notably, games and video games are used because they offer a suitable, self-contained simulation in which humans and AI can interact with each other. The aim is to model the internal and external influences of an AI and incorporate them into the training process to enable a time-dependent natural behavior of an AI. This makes it possible to model emotional states of an AI and to incorporate them into the AI behavior.
AI is seen as the new strong solution to many research problems. However, people often forget that AI is far from perfect and make naïve approaches to solve problems that "somehow" should be solved with AI. A more in-depth insight shows that many people overestimate what AI can do and underestimate how much data is actually needed. One assumption that is often wrongly made is that AI can "think" like a human being. If something is easy for a human, for example, a game like Tic-Tac-Toe, it cannot be hard for an AI, can it?
The way AIs are right now, they are mostly just programs that run what they were programmed to do. In this case, programming is a "training" on data, which, in the end, leads to the desired behavior. If we stick to the example of tic-tac-toe, we find that a neural network trained to win will always make the same move in the same situation. It is deterministic and therefore reacts differently than a human opponent would. This project is about teaching an AI to do something natural and, e.g., to have different emotional states when playing. A naturalAI bot can, e.g., play more aggressively or defensively, or sometimes do something unexpected. When people play with or against an AI, it should feel like playing with a human.