一种自适应强化学习算法在状态空间构建中的应用

Application of Adaptive Reinforcement Learning for State Space Construction

针对模型未知以及具有连续状态的系统控制问题,提出一种基于强化学习的自适应控制策略。在Actor-Critic框架下,建立归一化径向基网络的自适应调节机制,实现未知系统状态空间的动态创建。有效克服了状态空间分割所带来的维度灾难,而且能够使得系统的结构总保持在最佳状态。通过对倒立摆控制的仿真研究验证了方法的有效性。

In order to solve the control problem for unknown model system with continuous state, an adaptive control strategy based on reinforcement learning was proposed. Under the Actor-Critic architecture, the adaptive adjustment mechanism for normalized radial basis function network was established to realize the state space construction dynamically. This approach could overcome the curse of dimensionality caused by state space division effectively and make the system structure always stay the optimal status. Simulation research for inverted pendulum control demonstrates the validity of the proposed method.