| Abstract |
Inductive Logic Programming (ILP) is a form of symbolic machine learning, where the goal is to learn human- readable rules from examples. A key advantage of ILP systems over other forms of ML is that the rules they learn can be trans- lated into plain English, making them inherently explainable. However, many ILP systems do not scale over large domains, or to tasks with large numbers of examples or large search spaces. In our recent work, we developed a much more scalable ILP system called FastLAS. While the initial version of FastLAS is significantly faster than the previous state-of-the-art in ILP, it has some limitations which may reduce its usefulness in DAIS- related applications. For example, as DAIS scenarios tend to be dynamic, learning may need to be performed repeatedly as new data becomes available. In FastLAS, the learning process would need to be restarted from scratch each time, making learning at runtime infeasible.
In this paper, we discuss the current limitations of FastLAS and propose an improved ILP system, called FastLAS2, that overcomes each of these current limitations. In particular, we present the new features that make FastLAS2 scalable and effective even in uncertain domains. These include caching, which enables fast iterative learning; continuous data types, which allows it to be applied on datasets with continuous features; learning from partial examples, where some data may be uncertain or missing; non-observational predicate learning, which allows the learning of rules which are only indirectly linked to what we can observe. These features make FastLAS2 particularly suited to address policy-based ensemble learning, where both symbolic learning and statistical machine learning can be used, and neural-symbolic learning for learning human-readable rules from unstructured data in the presence of uncertainty.
|