Scholarlink

In the presence of exogenous events the Markov Decision Process (MDP) based state space representation of the academic exam timetable has exhibited limitations while generating schedules. That can be construed to flat state representations and slow policy learning via Value Iteration. In place of this, a relational graph-based representation with nodes are exam slots (referred to as states) and edges correspond to a decision assigning an exam to associated slot (stands for action) brings in non-linearity over relationships. The latent process of Graph Neural Network (GNN) enables message exchange among nodes leading to holistic structural understanding of the network by all nodes. Reinforcement Learning, specifically Deep Q-Learning estimates Q-values for each state-action pair. The exam schedule policy is derived at the GNN output layer. It is observed that by combining graph neural network and value iteration algorithm the exam schedule state space has been explored in the informed manner bringing higher rewards much earlier with policy convergence in 0.41s which is significantly faster than MDP based value iteration method’s 2 s.

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

Efficient Exams Scheduling in Uncertain Conditions: A Graph Neural Network and Deep Q-Learning for Value Iteration Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • D. Teja Santosh,
  • M. Raghava,
  • Jyothirmai Joshi,
  • Ch. V. S. Satyamurty,
  • Ch. Ram Mohan

摘要

In the presence of exogenous events the Markov Decision Process (MDP) based state space representation of the academic exam timetable has exhibited limitations while generating schedules. That can be construed to flat state representations and slow policy learning via Value Iteration. In place of this, a relational graph-based representation with nodes are exam slots (referred to as states) and edges correspond to a decision assigning an exam to associated slot (stands for action) brings in non-linearity over relationships. The latent process of Graph Neural Network (GNN) enables message exchange among nodes leading to holistic structural understanding of the network by all nodes. Reinforcement Learning, specifically Deep Q-Learning estimates Q-values for each state-action pair. The exam schedule policy is derived at the GNN output layer. It is observed that by combining graph neural network and value iteration algorithm the exam schedule state space has been explored in the informed manner bringing higher rewards much earlier with policy convergence in 0.41s which is significantly faster than MDP based value iteration method’s 2 s.