近年來廣告支持型內容平台，如YouTube 與 Bilibili，透過補貼政策激勵創作者提升內容品質，以吸引消費者觀看並提升平台廣告收益。本研究以深度強化學習(Deep Reinforcement Learning， DRL) 建構平台與創作者雙方為決策主體的 Stackelberg賽局架構，延伸Gaoetal.(2025) 的模型，並納入平台抽成比例與不同廣告邊際收益情境（低、基準、高），分析補貼策略與創作者行為之互動關係。透過DQN與PPO等深度強化學習方法求解策略均衡，並探討補貼策略選擇、利他消費者行為，以及抽成比例對平台與創作者利潤與內容品質決策的影響。 實驗結果顯示，在多種情境下均收斂到以流量補貼為主的策略，貨幣補貼僅 在特定廣告收益與抽成條件下出現，與既有理論結果一致。在策略行為分析中，當市場中利他消費者比例與利他程度同時達到一定門檻時，創作者傾向選擇自願付費模式。在抽成機制方面，兩種收費模式下平台利潤皆隨抽成比例提高而持續增加，而創作者利潤與內容品質則隨抽成比例提高而持續下降，顯示抽成比例的提高在促進平台收益的同時，亦會削弱創作者維持較高內容品質的誘因。不同演算法中，DQN、Double DQN，以及Dueling DQN 方法呈現較穩定之收斂結果，而PPO則展現較高探索性但收斂穩定性較低。;In recent years， advertising-supported content platforms such as YouTube and Bilibili have adopted subsidy policies to incentivize content creators to improve content quality， thereby attracting consumers and increasing advertising revenue. This study employs Deep Reinforcement Learning (DRL) to construct a Stackelberg game framework in which both the platform and content creators act as decision-making agents. Extending the model proposed by Gao et al. (2025)， this research incorporates platform commission rates and different advertising marginal revenue scenarios (low， baseline， and high) to investigate the interactions between platform subsidy strategies and creator behaviors. To solve the equilibrium strategies， several DRL algorithms， including Deep Q Network (DQN) and Proximal Policy Optimization (PPO)， are implemented. The study further examines the effects of subsidy strategy selection， altruistic consumer behavior， and platform commission rates on platform profits， creator profits， and content quality decisions.
The experimental results indicate that traffic-based subsidies emerge as the dominant strategy across most scenarios， while monetary subsidies appear only under specific combinations of advertising revenue and commission rates， consistent with existing theoretical findings. The behavioral analysis further reveals that creators tend to adopt a voluntary-pay pricing model when both the proportion of altruistic consumers and the degree of altruism exceed certain threshold levels. Regarding the commission mechanism， the equilibrium points of platform and creator profits under fixed-price and voluntary-pay settings are observed within different commission-rate ranges. Beyond these ranges， creator profits and investments in content quality gradually decline. Among the evaluated algorithms， DQN， Double DQN， and Dueling DQN demonstrate more stable convergence performance， whereas PPO exhibits greater exploration capability but lower convergence stability.
