Scholarlink

The article considers the problem of planning a portfolio of knowledge-intensive projects under conditions of limited resources, high uncertainty, and many conflicting performance criteria. Traditional optimization methods are hampered by the problem’s complexity, NP-hardness, and the lack of statistical data for forecasting. In this regard, a new approach based on reinforcement learning (RL) is proposed to construct near-optimal, rational plans for project implementation in polynomial time. The formalization of the problem includes taking into account time, financial, and capacity constraints, as well as project priorities and the sequence of work. The objective function maximizes the portfolio’s scientific and practical value, minimizes the waiting time for project start, and reduces costs for key resources. The RL environment is implemented using the stable_baselines3 library, which improves computational efficiency. Experimental results show a gradual improvement in solution quality during agent training, confirming the approach’s applicability to supporting decision-making in complex scientific and innovative projects. The proposed method offers opportunities to automate project portfolio management in high-dimensional, dynamic environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Using Reinforcement Learning Machine Models to Optimize Knowledge-Intensive Project Programs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Igor Kartsan,
  • Natalia Ovsyanikova,
  • Arutyun Khachaturyan

摘要

The article considers the problem of planning a portfolio of knowledge-intensive projects under conditions of limited resources, high uncertainty, and many conflicting performance criteria. Traditional optimization methods are hampered by the problem’s complexity, NP-hardness, and the lack of statistical data for forecasting. In this regard, a new approach based on reinforcement learning (RL) is proposed to construct near-optimal, rational plans for project implementation in polynomial time. The formalization of the problem includes taking into account time, financial, and capacity constraints, as well as project priorities and the sequence of work. The objective function maximizes the portfolio’s scientific and practical value, minimizes the waiting time for project start, and reduces costs for key resources. The RL environment is implemented using the stable_baselines3 library, which improves computational efficiency. Experimental results show a gradual improvement in solution quality during agent training, confirming the approach’s applicability to supporting decision-making in complex scientific and innovative projects. The proposed method offers opportunities to automate project portfolio management in high-dimensional, dynamic environments.