ABSTRACT
Least-squares conditional density estimation (LSCDE), introduced in Chapter 10, is a practical transition model estimator. However, transition model estimation is still challenging when the dimensionality of state and action spaces is high. In this chapter, a dimensionality reduction method is introduced to LSCDE which finds a low-dimensional expression of the original state and action vector that is relevant to predicting the next state. After mathematically formulating the problem of dimensionality reduction in Section 11.1, a detailed description of the dimensionality reduction algorithm based on squared-loss conditional entropy is provided in Section 11.2. Then numerical examples are given in Section 11.3, and this chapter is concluded in Section 11.4.
