Article: Dynamic optimisation of visitor diversion in smart scenic areas using deep reinforcement learning Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.70 pp.47 - 67 Abstract: Peak-time congestion in smart scenic areas often concentrates at a few hotspots and spreads quickly, raising safety pressure and degrading visitor experience. To address dynamic diversion under non-stationary demand, this paper proposes a constrained deep reinforcement learning framework for real-time guidance. First, a graph-based encoder captures spatial spillover among attractions and corridors. Then, a spatiotemporal attention module anticipates short-horizon surges and stabilises decisions. Finally, constraint-aware learning keeps recommendations within safety margins while balancing waiting, load equity, and throughput. Experiments on calibrated peak-demand scenarios show that the proposed method reduces average waiting time from 29.2 to 24.6 minutes and cuts safety violation rate from 2.0% to 1.2% compared with a vanilla learning baseline. relative to rule-based control, waiting drops from 38.6 to 24.6 minutes and near-violation time decreases from 41.2 to 14.8 minutes. The framework delivers robust improvements with steadier operating behaviour under diverse demand regimes. Inderscience Publishers - linking academia, business and industry through research

Title: Dynamic optimisation of visitor diversion in smart scenic areas using deep reinforcement learning

Authors: Qi Ma; Qinglin Wan

Addresses: Modern Business Travel College, Anhui Vocational College of City Management, Hefei, 230011, China ' Education Bureau of Hefei City, Hefei, 230071, China

Abstract: Peak-time congestion in smart scenic areas often concentrates at a few hotspots and spreads quickly, raising safety pressure and degrading visitor experience. To address dynamic diversion under non-stationary demand, this paper proposes a constrained deep reinforcement learning framework for real-time guidance. First, a graph-based encoder captures spatial spillover among attractions and corridors. Then, a spatiotemporal attention module anticipates short-horizon surges and stabilises decisions. Finally, constraint-aware learning keeps recommendations within safety margins while balancing waiting, load equity, and throughput. Experiments on calibrated peak-demand scenarios show that the proposed method reduces average waiting time from 29.2 to 24.6 minutes and cuts safety violation rate from 2.0% to 1.2% compared with a vanilla learning baseline. relative to rule-based control, waiting drops from 38.6 to 24.6 minutes and near-violation time decreases from 41.2 to 14.8 minutes. The framework delivers robust improvements with steadier operating behaviour under diverse demand regimes.

Keywords: smart scenic area; visitor diversion; crowd management; constrained deep reinforcement learning.

DOI: 10.1504/IJICT.2026.154384

International Journal of Information and Communication Technology, 2026 Vol.27 No.70, pp.47 - 67

Received: 25 Feb 2026
Accepted: 02 Apr 2026
Published online: 25 Jun 2026 *

Title: Dynamic optimisation of visitor diversion in smart scenic areas using deep reinforcement learning

Keep up-to-date