Optimizing high-dimensional stochastic forestry via reinforcement learning
研究用强化学习方法求解高维度的混合树种最优采伐模型，发现现有简化模型会损失大量经济产出或高估收益，而强化学习能高效处理年龄和大小结构。
In proceeding beyond the generic optimal rotation model, forest economic research has applied various specifications that aim to circumvent the problems of high dimensionality. We specify an age- and size-structured mixed-species optimal harvesting model with binary variables for harvest timing, stochastic stand growth, and stochastic prices. Reinforcement learning allows solving this high-dimensional model without simplifications. In addition to presenting new features in reservation price schedules and effects of stochasticity, our setup allows evaluating the simplifications in the existing research. We find that one- or two-dimensional models lose a high fraction of attainable economic output while the commonly applied size-structured matrix model overestimates economic profitability, yields deviations in harvest timing, including optimal rotation, and dilutes the effects of stochasticity. Reinforcement learning is found to be an efficient and promising method for detailed age- and size-structured optimization models in resource economics.