Abstract:
|
|
Operational constraint satisfaction in haul-truck dispatch planning is essential for ensuring safety, equipment longevity, and continuous
production at mine sites. Many operational constraints are non-Markovian, as they depend on execution history and anticipated future system evolution to ensure feasibility. In practice, such constraints are often handled outside the dispatch optimisation process
using heuristic controllers or human intervention. Monte Carlo Tree Search (MCTS) has emerged as an adaptable approximate approach for large-scale haul-truck dispatch optimisation. However, satisfying non-Markovian constraints within MCTS is challenging due
to incomplete tree exploration and uncertain future violation estimates. To address these challenges, this thesis investigates opportunity-cost-based and robustness-based violation measures, together with algorithms for balancing exploration, exploitation,
and constraint satisfaction in MCTS. The proposed non-Markovian constrained MCTS planners are evaluated in simulated mining domains of increasing complexity and constraint interdependence. Results demonstrate the potential of constrained MCTS methods for adaptable
non-Markovian constraint handling under dynamic operational conditions.
|
Bio:
|
|
Milan is a PhD researcher with the Rio Tinto Sydney Innovation Hub at the University of Sydney, undertaking her PhD in the School
of Aerospace, Mechanical and Mechatronic Engineering (AMME), working on constrained planning algorithms. She completed her MSc in Robotics at the University of Birmingham and her BTech in Electronics and Communication Engineering at National Institute of Technology
Calicut, India. Her research interests include task planning and decision-making under uncertainty. Her work spans applications in mining logistics, robotics, epidemiological modelling, and power signal analysis.
|
|
|
