On-line Tree Size Prediction using Incremental Models

It is often useful to know, in advance, how much effort will be required to solve a search problem. Current techniques for estimating the number of nodes that a bounded depth-first search will expand require copious amounts of off-line training and only make predictions in domains with integer-valued heuristic estimates. We present a new technique for estimating the number of nodes that a bounded depth-first search will expand. Our technique has two main benefits over previous approaches: 1) it will work in domains with real-valued heuristic estimates and 2) it can be trained on-line during an iteration of search. We show that our technique can predict as well as previous approaches when trained off-line. We also show that our new approach can be used on-line to accurately predict the number of nodes that an IDA* style search will encounter on subsequent iterations. We demonstrate the usefulness of this technique by controlling an IDA* search: using knowledge learned on previous iterations to set the bound for subsequent iterations. While our technique has more overhead than previous methods for controlling IDA*, it can give more robust performance by accurately doubling the amount of search effort between iterations.