ベクトル量子化による決定論的方策地図の不可逆圧縮 Lossy Compression of Deterministic Policy Map with Vector Quantization

For real-time decision making of a robot, there is an approach that utilizes the look-up table of the pre-computed result of dynamic programming. The look-up table records appropriate behavior for every situation of the robot and its surroundings. A robot that is installed the look-up table can decide its behavior only with a reference of the table. However, a table is usually too large to be loaded on the memory of usual robots. For the solution of this problem, we have proposed to use vector quantization for compressing the table. In this paper, we evaluate this method quantitatively. Then, we newly introduce an information entropy function that searches an appropriate way of blocking. For simulations and experiments, a look-up table for soccer behavior was created and compressed. As the results, the entropy function could find an appropriate way of blocking and the compression method with the blocking way enable the table to be compressed to 1.5% size.