Conference Agenda

Modelling and Optimization of end-to-end supply chain network which reflects daily planning and execution is challenging, especially when the network is complex with multi-echelon and the number of products within the network is large. In this paper, we propose a new method of modelling complex end-to-end supply chain network with both planning and execution activities, accelerating the simulation of network model with large number of products by decomposition, embedding business rules, and optimizing large amount of planning parameters using tailor designed optimization algorithm with embedded knowledge from business, advanced analytics and machine learning. The results demonstrate that the simulation and optimization process can be greatly speed up and the quality of proposed planning parameters can be well improved.

Farm Management Decisions, such as crop selection or, irrigation and fertilisation, are dependent on farmers’ discretion and may cause significant damage to the environment. This can be due to overuse of water and nitrogen-based fertilisers. Farmers may be forced to grow certain crops as they might be limited by available market (or supply chain) or existing economic policies. Economic policies, which are unilaterally prescriptive, may not include proper feedback to assess harm caused by selection of certain types of crops and their requirements. This may cause distress to farmers, loss to governments or supply chain entities, and damage to the environment. Hence, a comprehensive approach is needed for sustainability.

The proposed model consists of three levels:

1. Farms: Optimising the amount of ground water loss, nitrogen leaching, growth of crop main product and by-products. For this, we use gym-DSSAT that is a well-considered crop simulation model in a Reinforcement Learning environment.

2. Post-Harvest Processing: Optimising cost of processing, storing, or selling the processed products to distributors or end-consumers.

3. Distribution: Weekly capacity allocation and utilisation for product distribution.

Integration of above in a single model leads to high dimensionality. Therefore, Multi-Agent Hierarchical Reinforcement Learning has multiple advantages: (1) This is a modular approach for each level hence behaviours might be turned into specific tasks. (2) Since the high-level agent may ignore implementation specifics, exploration is improved. (3) Sample efficiency is improved. (4) This isolation of decisions leads to overall robustness. (5) Each policy can be transferred to new environments.

In this contribution, we address the challenge of anomaly detection within the logistics sector, especially supply chains, conceptualizing it as a task of categorizing logistics records into anomalous and normal categories. This classification is approached through both supervised and unsupervised learning methods. We detail the rationale behind selecting these methodologies and present a comparative analysis of their performance and efficiency, delving into the underlying reasons for their effectiveness.

The main findings indicate that for simpler logistic datasets, either supervised or unsupervised learning techniques can be effectively applied. However, for more complex datasets that necessitate guided learning, supervised methods become essential.

We introduce a novel unsupervised anomaly detection approach tailored to logistics, which leverages the strengths of autoencoders and clustering techniques to enhance detection speed and accuracy.

Additionally, leveraging the insights from feature importance metrics derived from supervised learning models, we further investigate the factors significantly impacting the timeliness of logistics operations. This exploration not only advances our understanding of anomaly detection in logistics but also contributes practical insights for enhancing operational efficiency in the domain.