Troubleshooting knowledge acquisition is a notorious network maintenance expert systems development bottleneck. We present an improved methodology to generate automatically a skeleton of network troubleshooting knowledge base given the data about network topology, test costs, and network component failure likelihood. Our methodology uses AO* search where a suitable modification of the Huffman code procedure is found to be an admissible heuristic. Our heuristic uses synergistically information about both component failure rates and test costs while relaxing topology constraints. The resulting expert system (XTAR) minimizes expected troubleshooting cost faster and learns better troubleshooting techniques during its operation.