Compact Representation of Knowledge Bases in ILP

Many inductive systems, including ILP systems, learn from a knowledge base that is structured around examples. In practical situations this example-centered representation can cause a lot of redundancy. For instance, when learning from episodes (e.g. from games), the knowledge base contains consecutive states of a world. Each state is usually described completely even though consecutive states may differ only slightly. Similar redundancies occur when the knowledge base stores examples that share common structures (e.g. when representing complex objects as machines or molecules). These two types of redundancies can place a heavy burden on memory resources. In this paper we propose a method for representing knowledge bases in a more efficient way. This is accomplished by building a graph that implicitly defines examples in terms of other structures. We evaluate our method in the context of learning a Go heuristic.status: publishe