ABSTRACT
This paper investigates how agents learn and perform efficient strategies by trying different actions in asymmetric confrontation setting. Firstly, we use Gomoku as an example to analyse the causes and impacts of asymmetric confrontation: the first mover gains higher power than the second mover. We find that the first mover learns how to attack quickly while it is difficult for the second mover to learn how to defend since it cannot win the first mover and always receives negative rewards. As such, the game is stuck at a deadlock in which the first mover cannot make further advances to learn how to defend, and the second mover learns nothing. Secondly, we propose an ancillary mechanism (AM) to add two principles to the agent’s actions to overcome this difficulty. AM is a guidance for the agents to reduce the learning difficulty and to improve their behavioural quality. To the best of our knowledge, this is the first study to define asymmetric confrontation in reinforcement learning and propose approaches to tackle such problems. In the numerical tests, we first conduct a simple human vs AI experiment to calibrate the learning process in asymmetric confrontation. Then, an experiment of 15*15 Gomoku game by letting two agents (with AM and without AM) compete is applied to check the potential of AM. Results show that adding AM can make both the first and the second movers become stronger in almost the same amount of calculation.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1. usually the most valuable one.
2. For example, the last node in the graph reached the value 1.
3. This process corresponds to step 1 in , because the value of the actual situation (parent node) is meaningless, so ignore it in the pattern.
4. positive reward for the winner, negative feedback for the loser.
5. assuming we conduct 400 MCTS searches, nearly half of the computing resource will be wasted on such useless wrong searches for the first layer of the board.
6. about 500 games.
7. In the search and the actual game, there will be deviations in the agent’s behaviour pattern. In order to ensure stability, 100 MCTS will be performed during the game, and the one with the most visits will be selected when the game is placed. When searching, especially the first exploration in this branch, the specific choice of grid placement depends entirely on the prior probability.
8. See Appendix.
9. For an agent that not well trained, its self-play winning rate is about 6:4 black:white.