A New Approach in Discrete Event System by Deep Monte Carlo Tree Search

The decision-making problem is one of core issues in discrete event system, which is often used to describe adversarial complex environment. Aiming at the problem that there exists huge space in this kind of environment, we use Monte Carlo Tree Search (MCTS) as our main search algorithm. One of the key challenges is that, during simulation, it is difficult to accurately predict opponent agents' actions. To deal with this challenge, we propose a deep neural network based on Long Short-Term Memory Network (LSTM), which is trained to predict most possible actions from opponent agents, the prediction is used to guide simulation. We test the performance of our method in the field of autonomous driving (AD), experiments are carried out on the simulation platform CARLA. We design a route containing several dangerous events. Results show that, if the prediction is accurate enough, algorithm using MCTS and deep learning outperforms the one using MCTS only, the former has a higher success rate in avoiding collisions, meanwhile it has a higher average speed.