Training with enlightening model for games with difficult-starting problem

Self-play process is known to work in training an AI Go player. But in some other games, such as Competition Mahjong, we find that only with this method it's hard to reach a usable final state for an agent of primary level, which makes training difficult in the early stages. This work presents a new learning framework, Training with Enlightening Model (TEM), that aims to solve the difficult-starting problem in games. We evaluate our framework with comparison to self-play model and show that it outperforms this baseline.