Applications of Curriculum Learning and Self-Play of Reinforcement Learning Techniques for Competitive Environments in Kazakh National Games

Reinforcement learning (RL) is a machine learning (ML) and deep learning technique that trains you to make decisions to achieve the most optimal results. This type of learning is based on a simulated trial-and-error learning process that is used to achieve specific goals. Actions aimed at achieving the goal are reinforced, and actions that distract from the goal are ignored. An intelligent entity or agent, interacting with the environment and receiving rewards or penalties for its actions, must make the most optimal decisions at each step to achieve its goal, for example, winning a game. The difficulty lies in building a reward system for the environment when training agents. Often, the rules of the game do not always specify the reward for each step of the agent in the environment, for example, in games such as chess, it is not easy to design the reward for the agents, since winning the game is the end result. The success of training directly depends on a well-structured and thoughtful reward strategy for the environment. Learning in multi-agent environments, i.e. where multiple agents (competitive or allied) interact is a challenging task because the agents are trained in parallel. To train several agents, the self-play technique is widely used, when agents learn by playing with each other. In this work, to train competitive agents, a multi-agent environment was written for the game “Togyzqumalaq”, for which it was possible to build a good reward system used in the agent training curriculum. Also, during training, the self-play technique made it possible to strengthen the agents’ skills to win the game.