Deep Learning (DL) techniques offer innovative solutions to automating DoD-relevant instruction. However, the improvement comes at the cost of large amounts of data. DL is not effective in data-sparse environments. Any application for DL without a dataset already available requires the laborious and expensive task of collecting data before outcomes can become useful. Within the DoD research community, operationally relevant datasets are difficult to acquire even when they have been collected, leading to difficulties applying DL techniques.
In spite of these challenges, we show that we can use domain knowledge and machine transfer learning to make initial progress while data is being collected. This paper presents a case study in transfer between random and published puzzles while using (DL) approaches to solve Sudoku. The Sudoku represent an instructional domain providing controlled evaluation of DL outputs. A published Sudoku puzzle uses the interplay of different patterns to engage a player, which represents a subset of all possible game boards. The patterns in published puzzles are analogous to the nuances within a domain that characterize operational data. Given the cost of operational data, we desire a DL approach that performs well with few published puzzles as training inputs.
We describe an approach that reduces the data requirement and increases the performance of DL when little data is available. The approach uses transfer from random training data to speed and enhance DL training on data with patterns reflecting operational characteristics. Using the Sudoku domain, we show that transfer from random generated puzzles makes DL efficient after relatively few published examples are added. Furthermore, DL performance increases over time as real data is added. The benefits are hypothesized to support delivering instruction in settings with new and emerging tactics.