, we propose a convolutional neural network-based virtual

The proper application of retrosynthesis to identify possible transformations for a given target compound requires a lot of chemistry knowledge and experience. However, because the complexity of this technique scales together with the complexity of the target, efficient application on compounds with intricate molecular structures becomes almost impossible for human chemists. The idea of using computers in such situations has existed for a long time, but the accuracy was not sufficient for practical applications. Nevertheless, with the steady improvement of machine learning and artificial intelligence in recent years, computer-assisted retrosynthesis has been gaining research attention again. Because of the overall lack of chemical reaction data, the main challenge for the recent retrosynthesis methods is low exploration ability during the analysis of target and intermediate compounds. The main goal of this research is to develop a novel, template-free approach to address this issue. Only individual molecular substructures of the target are used to determine potential disconnection sites, without relying on additional information such as chemical reaction class. The model for the identification of potential disconnection sites is trained on novel molecular substructure fingerprint representations. For each of the disconnections suggested using the model, a simple structural similarity-based reactant retrieval and scoring method is applied, and the suggestions are completed. This method achieves 47.2% top-1 accuracy for the single-step retrosynthesis task on the processed United States Patent Office dataset. Furthermore, if the predicted reaction class is used to narrow down the reactant candidate search space, the performance is improved to 61.4% top-1 accuracy.

In all information systems it is very important to operate with correct information. Incorrect information can lead to many problems that can cause direct financial and reputation loss of the company. Data used by the system can be gathered by sensors, scripts or by hand. In all those cases, mistakes are possible. It is important to detect mistakes on time and stop them from propagating further into the system. In this paper, a novel multi-step anomaly detection algorithm based on the greatest common divisor and median value is described. The algorithm for anomaly detection in historical sales data is used as a part of the smart warehouse management system which is implemented in some of the largest distribution companies in Bosnia and Herzegovina. The algorithm showed significant results in anomaly detection on company orders and improved a number of processes in the operation of the smart warehouse management system. The algorithm described can also be used in other areas where the transaction data is collected, such as sales and banking,

One of the frequently occurring tasks during the development of a warehouse management system is the implementation of a routing algorithm of some kind. Whether it is for guiding workers during order picking, routing delivery vehicles or for routing company representatives, this task has proven to be challenging in the technical as well as the social sense. In other words, the task is heavily dependent on various company-specific constraints and it directly dictates the way employees should do their job. This paper describes a strategic approach to the development and gradual integration process of such algorithms which makes sure that all constraints are satisfied and, more importantly, ensures that route suggestions are viewed by the employees as a helpful tool rather than a threat to their job. Described through a real-world case study in a medium-to-large warehouse, the routing efficiency is almost doubled in comparison to the previous approach and critical factors are analysed and discussed throughout different stages of the process.

One of the frequently occurring tasks during the development of warehouse management systems is the implementation of routing algorithms of some kind. Whether it is for routing workers during order picking, delivery vehicles or company representatives, this task has proven to be challenging in the technical as well as the social sense. In other words, the task is heavily dependent on various general and company-specific constraints and it directly dictates the way employees should do their job. This paper describes a strategic approach to the development and gradual integration of such algorithms which makes sure that all constraints are satisfied and, more importantly, ensures that route suggestions are viewed by the employees as a helpful tool rather than a threat to their job. In the first part of this paper, the approach is described and evaluated on a warehouse representative routing problem through a real-world case study in a medium-to-large warehouse. In the second part, the same approach is adapted to a delivery vehicle routing problem for a smaller retailer company. In both cases, routing efficiency almost doubled in comparison to previous approaches used by the companies. The most important factors of the implementation and integration stages as well as the impact of the changes on employee satisfaction are aggregated, analysed in detail, and discussed throughout different stages of development.

In modern market it is very important to deliver products to customers fast. That delivery can be on site or to customer's homes. In order to achieve that it is important to have enough goods stored in warehouses and prepared for delivery. It is not a good decision to clutter up warehouses with the goods because space is limited and expensive and it makes it more complicated to collect orders. Those are the reasons why it is important that number of stored goods converge to the exact number of product units that will be ordered in the future. Demand forecasting tries to solve that problem. In this work demand forecasting algorithm based on Long Short-Term Memory recurrent neural network is described and compared with demand forecasting algorithms developed by authors before.

Building a successful warehouse management system encompasses solving many problems of different nature to reshape the general workflow and ensure improvements in terms of resource management. In order for such a system to be accepted and used by a logistics company, those solutions need to be presented through a simple, adaptable and most importantly, a feasible software solution. One of the aspects that needs to be covered while building a warehouse management system is the optimal product placement in the warehouse. If the products are strategically placed, all the other improvement strategies like stock to picking zone item transfer and item picking order become more efficient and easier to implement. In this paper, tactics and issues regarding the optimal product placement in a warehouse are analyzed in detail through a real-world case study. Solutions of this problem largely differ for new, empty warehouses and for already operative warehouses not able to carry out a stocktaking process. Approaches for both possible situations are proposed and tested out on two different warehouses in a medium-to-large logistics company.

The warehouse layout directly affects the process of receiving and storing of goods, as well as the order picking process. A standard layout has its pallet places arranged in parallel shelves with aligned cross-aisles. However, a big number of warehouses use a shelf layout in which the complete warehouse cannot be represented as a sequence of parallel shelves. Effective calculation of distances between positions in the warehouse presents a significant step towards the effective order collection and distribution of goods. This paper describes a generic approach for calculating distances within a warehouse which does not necessarily have standard layout. It describes the application of the algorithm on different warehouse layouts in which the shelves can be split into smaller units so that inside of each unit the shelves are distributed in a standard way. Dynamic programming was used for the calculation of distances inside of those units. An analysis and testing of the algorithm were performed on two middle-sized warehouses with non-standard shelf layout. The algorithm was tested in the process of collecting orders, as well as the process of moving goods from the stock to the pick zone.

Efficient detection of distresses on asphalt pavements has a great impact on safe driving, thus it has been very active research subject in recent years. High severity level distresses, such as potholes, are the most severe threat to safe driving, hence timely detecting and repairing potholes is crucial in ensuring safety and quality of driving. Existing methods often require sophisticated equipment and algorithms with high-computational pre-processing steps for analysis of substantial amount of existing data (images or videos). In this paper, a new unsupervised method for detection of high severity distresses on asphalt pavements was proposed. The method was tested on highly unstructured image data set captured from different cameras and angles, with different irregular shapes and number of potholes to demonstrate its capability. Results indicated that the method can be used for rough detection and estimation of damaged pavements.