Higher order Q-Learning

Higher order learning is a statistical relational learning framework in which relationships between different instances of the same class are leveraged (Ganiz, Lytkin and Pottenger, 2009). Learning can be supervised or unsupervised. In contrast, reinforcement learning (Q-Learning) is a technique for learning in an unknown state space. Action selection is often based on a greedy, or epsilon greedy approach. The problem with this approach is that there is often a large amount of initial exploration before convergence. In this article we introduce a novel approach to this problem that treats a state space as a collection of data from which latent information can be extrapolated. From this data, we classify actions as leading to a high reward or low reward, and formulate behaviors based on this information. We provide experimental evidence that this technique drastically reduces the amount of exploration required in the initial stages of learning. We evaluate our algorithm in a well-known reinforcement learning domain, grid-world.