Faster conceptual blending predictors on relational time series

Tasks at upper levels of sensor fusion are usually concerned with situation or impact assessment, which might consist of predictions of future events. Very often, the identity and relations of target of interest have already been established, and can be represented as relational data. Hence, we can expect a stream of relational data arriving at our agent input as the situation updates. The prediction task can then be expressed as a function of this stream of relational data. Run-time learning to predict a stream of percepts in an unknown and possibly complex environment is a hard problem, and especially so when a serious attempt needs to be made even on the first few percepts. When the percepts are relational (logical atoms), the most common practical technologies require engineering by a human expert and so are not applicable. We briefly describe and compare several approaches which do not have this requirement on the initial hundred percepts of a benchmark domain. The most promising approach extends existing approaches by a partial matching algorithm inspired by theory of conceptual blending. This technique enables predictions in novel situations where the original approach fails, and significantly improves prediction performance overall. However an implementation, based on backtracking, may be too slow for many implementations. We provide an accelerated approximate algorithm based on best-first and A* search, which is much faster than the initial implementation.