During reading some papers about fuzzy systems, I've recognized a subtopic called “fuzzy commands”. The idea is to provide a sentence in natural language, like “move ball to right”, and the fuzzy reinforcement learning algorithm will determine by it's own how to fulfill the task. That means the fuzzy rules are adapted, they are learned.
From other papers about reinforcement learning it is known, that such systems are able to adapt to new problems. That means, the same reinforcement learning algorithm can be used to solve the cart-balancing problem as well as the pong game or the inverted pendulum problem. This is usually done with policy iteration. The policy is stored in a q-table and describes state-action values or it is described by a markov chain which is probabilistic transition diagram. If this is combined with natural language what is the result? Is a fuzzy command reinforcement learning system equal to a general game playing agent? Or has such a system some kind of disadvantages, that means it fails to solve more complex problems?