Learning & Prediction in Relational Time Series: A Survey

Making sense out of a stream of incoming percepts is the first step in any agent's cognition process. The purpose of sense-making is usually to facilitate sound decision making, often by making predictions of future events or actions. In the case that the percepts are relational, the technologies available for this task are mainly based on production systems or statistical graphical model inferencing processes such as Bayesian networks. To apply these approaches, it is necessary that domain knowledge be known or that examples are available to a supervised learning process. Darken (2005) proposed a situation learning (SL) approach to learn a string of percept sequence into a set of overlapping situations. This approach has much potential for learning and predicting in domains that are characterized by high variability and great number of predicates and terms that become known only at runtime, and which feature a trending or moving context environment. In this paper, we attempt to define relational time series (RTS) and its characteristics for evaluating current learning approaches for learning and prediction of RTS. We also report the prediction accuracies of various prediction techniques based on SL in a benchmark environment.