Abstract:

Reinforcement learning is often used to dynamically improve strategy of solving logic problems on varying complexity levels. In a number of game problems machine learning algorithms already surpassed top level human capabilities. As artificial neural network algorithms and hardware computational capabilities quickly improve, more complex problems can be solved. One of the computational challenges is growing number of decisive agents. Each decisive agent consume resources and it might be viable to execute simulation in distributed environment. Paper presents conception and actual implementation state of type and memory safe simulation library with respect to reinforcement learning requirements. Briefly the indications leading to selecting Rust as implementation language are explained. Later discussed are restrictions and freedoms regarding asymmetric agent execution and network communication. Paper contains also description of problems with flexibility of dynamic model under type and memory safety restrictions of Rust language. At the end of the paper, a direction of future development is indicated.