Choose timezone
Your profile timezone:
Powered by Indico
v3.3.3-pre
Internet intrusion detection systems (IDS) use machine learning models, which one needs to train using public datasets. The training process requires a training set, which is a majority part of such a dataset, while validation is performed on its second part - the validation set. Finally, to evaluate the quality of the output model, one utilizes the test set, which is the third part. The measure (accuracy, precision, recall, or F1) obtained on the latter determines the quality of the model. In our paper, we investigate how the model prepared in the above classic way performs on other data, i.e., to what extent the model is biased to the public dataset used in the IDS model preparation. Our investigation uses cross-validation of models based on four internet traffic datasets: UNSW-NB15, BoT-IoT, ToN-IoT, and CIC-CSE-IDS2018. The results obtained show that quality measures of a model trained on one public dataset are only partially repeatable on others. It confirms the necessity of careful selection of data used in the machine learning models in IDS that guarantee high data diversity.
This paper proposes a solution to the problem of automatically generating an efficient front-end layer code based on a provided UI design. With the increasing complexity and scalability requirements of systems, companies, as well as individual developers, are inclined to seek a sufficient and maintainable way to automate some internal processes. One such automation may involve focusing on creating a visual prototype of a graphical user interface, rather than programming skills necessary to implement it. This approach makes full automation achievable in cases where the front-end layer is a secondary concern.
The research described in the paper utilizes Node.js scripts paired with React utilities to transform a hierarchical JSON description of UI design into a basic React application. It then injects it into JSX templates while maintaining the same hierarchy in the generated virtual DOM as provided. The aspect of efficiency is addressed by providing a way to group similar components and reuse them. The resulting application is compared with referential solutions in terms of TypeScript rules compliance, code complexity, and memory allocation during the execution of a simple use case scenario.
keywords: front-end, code generation, UI design, React
The paper explores different solutions for implementing self-learning artificial intelligence (AI) competitive bots for the game Blood Bowl. The winners of the most of the previous competitions were scripted bots but in recent years bots based on machine learning started to outspace their competition. Blood Bowl is a two-player, turn-based, asymmetric board game that combines elements of American football with the Warhammer board game. Teams consist of eleven to sixteen players, each of them having varying configurations of five main statistics: move allowance, strength, agility, armor value, and passing. The main goal is to score a higher takedown number than the opponent. This paper’s primary objective is to develop a sophisticated AI agent capable of participating in the Bot Bowl Tournament and competing against other state-of-the-art bots. The research focuses on exploring behavioral cloning solutions created by using an in-depth analysis of games played in previous tournaments to vastly improve both the win ratio and complexity of moves employed by the bot. Using wrappers and scripted actions enhances the efficiency and effectiveness of the AI's learning mechanisms. By leveraging insights acquired from past gameplay data and employing advanced machine learning techniques, this research seeks to contribute to the advancement of AI in competitive gaming environments.
The paper discusses the significance of one-shot learning from prototype SKU images for efficient product recognition in various retail and inventory management sectors. Traditional methods require large supervised datasets for training deep neural networks, which can be costly and impractical. One-shot learning techniques address this issue by enabling classification from a single prototype image per product class, reducing data annotation efforts. The variational prototyping-encoder (VPE) is introduced as a novel deep neural network tailored for one-shot classification. By utilizing a support set of prototype SKU images, VPE learns to classify query images while capturing image similarity and prototypical concepts. Unlike metric learning-based approaches, VPE pre-learns image translation from real-world object images to prototype images as a meta-task, facilitating efficient one-shot classification with minimal supervision. The result of the research indicated the potential for applicability of VPE in reducing the need for large datasets and accurately classifying query images into their respective categories, offering a practical solution for product classification tasks.
Index Terms: one-shot learning, autoencoders, VPE, prototyping.
Procedural plot generation is a topic widely researched in the context of video games. This paper discusses parts of the existing research using Role Playing Games as a target for plot generation. Analysis of table-, atom-, and Large Language Model-based approaches to plot generation for Role Playing Games indicates that more is needed. This paper proposes the solution to this problem using a Polish Role Playing Game named Wolsung as an example system not well known by the Large Language Model. Using tables, a skeleton is created and responsible for maintaining word information and keeping causality. The resulting skeleton is then transformed with the GPT model to convert the skeleton into an actual plot. Evaluation of results is performed based on language correctness and task completion. The solution provides a way of using the Large Language Model over a broad unknown domain without the need for additional training.