System identification, also known as learning forward models, transfer functions, system dynamics, etc., has a long tradition both in science and engineering in different fields. Particularly, it is a recurring theme in Reinforcement Learning research, where forward models approximate the state transition function of a Markov Decision Process by learning a mapping function from current state and action to the next state. This problem is commonly defined as a Supervised Learning problem in a direct way. This common approach faces several difficulties due to the inherent complexities of the dynamics to learn, for example, delayed effects, high non-linearity, non-stationarity, partial observability and, more important, error accumulation when using bootstrapped predictions (predictions based on past predictions), over large time horizons. Here we explore the use of Reinforcement Learning in this problem. We elaborate on why and how this problem fits naturally and sound as a Reinforcement Learning problem, and present some experimental results that demonstrate RL is a promising technique to solve these kind of problems.
In a Role-Playing Game, finding optimal trajectories is one of the most important tasks. In fact, the strategy decision system becomes a key component of a game engine. Determining the way in which decisions are taken (online, batch or simulated) and the consumed resources in decision making (e.g. execution time, memory) will influence, in mayor degree, the game performance. When classical search algorithms such as A* can be used, they are the very first option. Nevertheless, such methods rely on precise and complete models of the search space, and there are many interesting scenarios where their application is not possible. Then, model free methods for sequential decision making under uncertainty are the best choice. In this paper, we propose a heuristic planning strategy to incorporate the ability of heuristic-search in path-finding into a Dyna agent. The proposed Dyna-H algorithm, as A* does, selects branches more likely to produce outcomes than other branches. Besides, it has the advantages of being a model-free online reinforcement learning algorithm. The proposal was evaluated against the one-step Q-Learning and Dyna-Q algorithms obtaining excellent experimental results: Dyna-H significantly overcomes both methods in all experiments. We suggest also, a functional analogy between the proposed sampling from worst trajectories heuristic and the role of dreams (e.g. nightmares) in human behavior.