Q-concept-learning: generalization with concept lattice representation in reinforcement learning
Marc Ricordeau · 2004
One of the very interesting properties of reinforcement learning algorithms is that they allow learning without prior knowledge of the environment. However, when the agents use algorithms that enable a generalization of the learning, they are unable to explain their choices. Neural networks are good examples of this problem. After a reminder about the basis of reinforcement learning, the lattice concept will be introduced. Then, Q-concept-learning, a reinforcement learning algorithm that enables a generalization of the learning, the use of structured languages as well as an explanation of the agent's choices will be presented.