Esports is growing rapidly, yet the data available to researchers is limited due to the game company policies. Consequently, vision-based approaches utilizing game screens are gaining attention as a practical alternative. We focus on the League of Legends minimap and address the challenges of champion detection when extracting champion information from the minimap. The challenges in this domain include small objects, rapid movement, and frequent occlusions. We propose a transfer-learning-based object detection pipeline that combines synthetic data with a subset of replay data. Synthetic data enables the rapid generation of diverse scenarios and improves training scalability, while replay data reduces the data distribution gap. This approach achieves 0.588 mean average precision, improving over replay-only by 0.261 and synthetic-only by 0.312, with 6.4 ms latency. Furthermore, we constructed a dataset encompassing all champions, enabling comparative analysis of detection models and supporting reproducible benchmarking for various application studies.

Instrumental music conveys rich affective experiences through acoustic cues, yet instrumental passages often remain inaccessible to Deaf and Hard-of-Hearing (DHH) audiences. Although captioning practices for vocal songs have expanded, instrumental music remains largely uncaptioned, with no established criteria for representing musical content in text. We propose µCap (Music Captions), an automatic instrumental music captioning system that transforms instrumental audio into time-aligned, non-lexical textual renderings enhanced with simple visuals. Drawing on Preliminary surveys with DHH individuals and expert group discussions, we developed a phonetic-like captioning schema grounded in music sound analysis and linguistics. We then implemented µCap using audio feature extraction and a retrieval-augmented generation pipeline to produce expressive, sound-mimetic captions. Two user evaluations with DHH participants (n=20 and n=15) showed that µCap enhanced music appreciation, immersion, and perceived presence of acoustic detail. This work contributes empirical evidence and insights for designing caption-based visual representations that make instrumental music more accessible.

Self-annotation is the gold standard for collecting affective state labels in affective computing. Existing methods typically rely on full annotation, requiring users to continuously label affective states across entire sessions. While this process yields fine-grained data, it is time-consuming, cognitively demanding, and prone to fatigue and errors. To address these issues, we present PREFAB, a low-budget retrospective self-annotation method that targets affective inflection regions rather than full annotation. Grounded in the peak-end rule and ordinal representations of emotion, PREFAB employs a preference learning model to detect relative affective changes, directing annotators to label only selected segments while interpolating the remainder of the stimulus. We further introduce a preview mechanism that provides brief contextual cues to assist annotation. We evaluate PREFAB through a technical performance study and a 25-participant user study. Results show that PREFAB outperforms baselines in modeling affective inflections while mitigating workload (and conditionally mitigating temporal burden). Importantly, PREFAB improves annotator confidence without degrading annotation quality.

Although students need to self-monitor and manage their learning process for effective metacognition, it can be particularly challenging in solitary e-learning environments that rely on pre-recorded videos. Unlike interactive e-learning or physical classrooms, typical e-learning environments prevent students from interacting with their teachers and peers, thereby hindering metacognitive support. To address this challenge, we introduce GPTalk, a system designed to support students’ learning experiences by facilitating interactions with LLM-based virtual companions. Through interviews with students and teachers, we identified design recommendations and implemented them in GPTalk. A user study involving 32 high-school students demonstrated that, compared to a baseline system, GPTalk fostered richer metacognitive engagement and self-regulated learning processes during video-based study (e.g., more monitoring questions and in-situ reflections), while short-term content understanding accuracy remained comparable across conditions. Overall, our findings suggest that students’ interactions with a virtual teacher and peer can support key aspects of their metacognition and self-regulated e-learning processes.

Recent research has highlighted the significance of natural language in enhancing the controllability of generative models. While various efforts have been made to leverage natural language for content generation, research on deep reinforcement learning (DRL) agents utilizing text-based instructions for procedural content generation remains limited. In this paper, we propose IPCGRL, an instruction-based procedural content generation method via reinforcement learning, which incorporates a sentence embedding model. IPCGRL fine-tunes task-specific embedding representations to effectively compress game-level conditions. We evaluate IPCGRL in a two-dimensional level generation task and compare its performance with a general-purpose embedding method. The results indicate that IPCGRL achieves up to a 21.4 % improvement in controllability and a 17.2 % improvement in generalizability for unseen instructions with varied condition expressions within the same task. Furthermore, the proposed method extends the modality of conditional input, enabling a more flexible and expressive interaction framework for procedural content generation.

In esports broadcasting, human observers are tasked with providing viewers with a satisfying view of the event. Existing approaches focus primarily on detecting events and often fail to address how the viewing camera should transition after an event is detected. In this study, we defined the viewing process as event detection and viewport tracking, which is the process of following the detected event. In addition, we focus more on viewport tracking and propose a ConvLSTM-based encoder-decoder model based on the viewing data of professional observers. The model aims to automatically predict how observers follow events by learning their temporal patterns and spatial characteristics, and to build a more natural and effective esports automatic broadcasting system. The viewport tracking model was evaluated using Intersection over Union(IoU) and achieved a performance of 0.6689. This result represents a novel attempt to model the viewing sequence after event detection, which offers the potential to enhance the naturalness of automated broadcasting systems.

Plasticity loss, a critical challenge in neural network training, limits a model’s ability to adapt to new tasks or shifts in data distribution. While widely used techniques like L2 regularization and Layer Normalization have proven effective in mitigating this issue, Dropout remains notably ineffective. This paper introduces AID (Activation by Interval-wise Dropout), a novel method inspired by Dropout, designed to address plasticity loss. Unlike Dropout, AID generates subnetworks by applying Dropout with different probabilities on each preactivation interval. Theoretical analysis reveals that AID regularizes the network, promoting behavior analogous to that of deep linear networks, which do not suffer from plasticity loss. We validate the effectiveness of AID in maintaining plasticity across various benchmarks, including continual learning tasks on standard image classification datasets such as CIFAR10, CIFAR100, and TinyImageNet. Furthermore, we show that AID enhances reinforcement learning performance in the Arcade Learning Environment benchmark.

In this paper, we present an adaptive walker system designed to address limitations in current intelligent walker technologies. While recent advancements have been made in this field, existing systems often struggle to seamlessly interpret user intent for speed control and lack adaptability across diverse scenarios and terrain. Our proposed solution incorporates high-resolution tactile sensors, deep learning algorithms, IMU sensors, and linear motors to dynamically adjust to the user's intentions and terrain changes. The system is capable of predicting the user's desired speed with an error margin of only 20.99%, relying solely on tactile input from hand and arm contact points. Additionally, it maintains the walker's horizontal stability with an error of less than 1 degree by adjusting leg lengths in response to variations in ground angle. This adaptive walker enhances user safety and comfort, particularly for individuals with reduced strength or cognitive abilities, and offers reliable assistance on uneven terrain such as uphill and downhill paths.

In the domain of game artificial intelligence, tutorial systems have seen limited advancement despite their critical role in onboarding players. Traditional tutorials often neglect individual learning differences, leading to ineffective instruction. This study proposes a personalized in-game tutorial generation framework that leverages procedural content generation (PCG) and student modeling. The system integrates a Monte Carlo Tree Search (MCTS)-based state transition generator and a player modeling module to dynamically adapt tutorial content based on inferred internal knowledge. The approach is validated through large-scale user testing ($N$=88) in a commercial-style match-3 puzzle game environment. Results show that the personalized generator improves learning by up to 44.4% within short sessions. The findings highlight the potential of adaptive tutorials in accelerating skill acquisition and enhancing game experience through seamless, personalized learning contexts.

Game music critically influences the experience of a video game. Although this influence has been well investigated, the multifaceted relationships between video games and the emotions evoked by music are rarely reported. By considering diverse emotional matches of game and music, game designers could enhance various aspects of the game experience. The present study investigates players' game experiences by analyzing the electroencephalogram data, game-experience questionnaire answers, and interview responses of 31 experimental participants corresponding to game–music emotional matching based on the valence–arousal model. Finally, four findings were identified based on four types of game experiences: overall preference, emotion, immersion, and performance. These findings led to four game music design approaches.

Driven by the rapid growth of machine learning, recent advances in game artificial intelligence (AI) have significantly impacted productivity across various gaming genres. Reward design plays a pivotal role in training game AI models, wherein researchers implement concepts of specific reward functions. However, despite the presence of AI, the reward design process predominantly remains in the domain of human experts, as it is heavily reliant on their creativity and engineering skills. Therefore, this paper proposes ChatPCG, a large language model (LLM)-driven reward design framework. It leverages humanlevel insights, coupled with game expertise, to generate rewards tailored to specific game features automatically. Moreover, ChatPCG is integrated with deep reinforcement learning, demonstrating its potential for multiplayer game content generation tasks. The results suggest that the proposed LLM exhibits the capability to comprehend game mechanics and content generation tasks, enabling tailored content generation for a specified game. This study not only highlights the potential for improving accessibility in content generation but also aims to streamline the game AI development process.

The game Angry Birds is a challenging problem for artificial intelligence in that it requires physical reasoning ability. Previous approaches require domain knowledge or playing data, or have limitations in generalization performance. Inspired by human approach to physics-based puzzle games, we devise a new Angry Birds agent that separates AI’s thought into two stages. To this end, our method, LangBirds, uses Large Language Models (LLMs) that are recently considered to show human-level performance. We compared LangBirds to several reinforcement learning agents as well as heuristic agents. The results on the Phy-Q benchmark, which is a testbed based on Angry Birds, showed that our approach outperforms baselines. Moreover, the proposed approach allows us to understand the decision-making process since it uses natural language. Qualitative assessments indicated that the rationale for the decisions was reasonable.

Locomotion gesture classification in virtual reality (VR) is the process of analyzing and identifying specific user movements in the real world to navigate virtual environments. However, existing methods often necessitate the use of wearable sensors, which present limitations. To address this, we utilize a high-resolution carpet-type tactile sensor as a foot action recognition interface, which was previously unexplored in the context of locomotion gesture classification. This interface can capture the user’s foot pressure data in detail to distinguish similar actions. In this paper, to efficiently process captured user’s foot tactile data and classify nuanced actions, we utilize a Vision Transformer (ViT) architecture and propose a novel Self-Teaching Vision Transformer (STViT) model integrating elements of the Shifted window Vision Transformer (SwinViT) and Data-efficient image Transformer (DeiT). However, unlike DeiT, our model uses itself from $N -$steps prior as the teacher model, which is continuously updated. Therefore, improving the ability to classify actions by referencing its own knowledge from previous training stages progressively refines its understanding of similar action gestures. Also, we used the base architecture of SwinViT to utilize patch merging, which improves the ability to differentiate between variations in similar actions by capturing information at different scales. We evaluated seven vision-based methods, demonstrating promising results. Not only did our model outperform ResNet by 19.6%, but it also outperformed each Deit and SwinViT by 3.3% and 2.9%, achieving 92.7% accuracy. To validate our model’s real-world applicability, we conducted user preference tests and in-game performance evaluations with 18 participants. As a result, the participants preferred our model to SwinViT and DeiT, backing up the computational results. The video demonstrating the VR testing for STViT can be found in https://youtu.be/NJslvanRn18

The balance of game content significantly impacts the gaming experience. Unbalanced game content diminishes engagement or increases frustration because of repetitive failure. Although game designers intend to adjust the difficulty of game content, this is a repetitive, labor-intensive, and challenging process, especially for commercial-level games with extensive content. To address this issue, the game research community has explored automated game balancing using artificial intelligence (AI) techniques. However, previous studies have focused on limited game content and did not consider the importance of the generalization ability of play-testing agents when encountering content changes. In this study, we propose RaidEnv, a new game simulator that includes diverse and customizable content for the boss raid scenario in the MMORPG games. Additionally, we design two benchmarks for the boss raid scenario that can aid in the practical application of game AI. These benchmarks address two open problems in automatic content balancing, and we introduce two evaluation metrics to provide guidance for AI in automatic content balancing. This novel game research platform expands the frontiers of automatic game balancing problems and offers a framework within a realistic game production pipeline. The open-source environment is available at a GitHub repository.

Multiplayer contents play an essential role in extending the lifespan of the game and positively affect the player’s experience. However, previous studies on procedural content generation focused on single-player games and few for multiplayer games. Multiplayer games have different ethics and mechanics associated with competition or cooperation. Thus, content designers are concerned about this interaction. In this paper, we propose a new method for generating multiplayer levels that encourage diverse cooperation among game players. Our contributions are summarized as follows: 1) We designed four cooperation patterns usable in controllable generation literature. 2) We proposed a controllable level generator to deploy the proposed patterns in the two-player cooperative cooking game, Overcooked!. 3) We discussed effective methods for leading players’ cooperation experience. Consequently, we found that the players’ interaction is most fundamental at the multiplayer level. The results of our study will lead to diverse cooperation experiences in future multiplayer game content generation studies.

Recently, several attempts have been made to train cooperative artificial intelligence (AI). From training superhuman-level agents to human-like agents, the purpose of an AI results in differences in the behavior policy. Indeed, training a human-like agent could enhance the experience of multiplayer game players. However, training human-like agents is challenging and there is little existing work concerning benchmarking cooperative agents with actual humans. As an initial step to address this problem, we suggest a software program and an experimental procedure to conduct Turing tests in multiplayer games. Our contribution will help current multiagent studies benchmark the human-likeness of the agents and investigate their characteristics.

Humans are usually good at guessing whether the two games are similar to each other and easily estimate how much time to master new games based on the similarity. Although Deep Reinforcement Learning (DRL) has been successful in various domains, it takes much training time to get a successful controller for a single game. Therefore, there has been much demand for the use of transfer learning to speed up reinforcement learning across multiple tasks. If we can automatically determine the possibility of transfer learning in DRL domain before training, it could efficiently transfer knowledge across multiple games. In this work, we propose a simple testing method, Determining the Possibility of Transfer Learning (DPTL), to determine the transferability of models based on Grad-CAM visualization of the CNN layer from the source model. Experimental results on Atari games show that the transferability measure is successfully suggesting the possibility of transfer learning.

Deep Reinforcement Learning (DRL) allows agents to learn strategies to solve complex tasks. It has been applied to solve various problems such as natural language processing, games, etc. However, it is still difficult to apply DRL to certain real-world problems because each action is not predictable, and we cannot know why the results are coming out. For this reason, a technology called eXplainable Artificial Intelligence (XAI) has been recently developed. As this technology shows a visualization of the AI process, people can easily understand the results of AI. In this paper, we proposed to use Grad-CAM, one of the XAI techniques, when we visualize the behaviors of AI players trained by DRL. Our experimental results show which part of the input state is focused on when one well-trained agent takes action.

Real-time strategy games have become an increasingly popular test bed for modern artificial intelligence (AI) techniques. With this rise in popularity has come the creation of several annual competitions, in which AI agents (bots) play the full game of StarCraft: Broodwar by Blizzard Entertainment. The three major annual StarCraft AI Competitions are the Student StarCraft AI Tournament, the Computational Intelligence in Games competition, and the Artificial Intelligence and Interactive Digital Entertainment competition. In this paper, we will give an overview of the current state of these competitions, describe the bots that compete in them, and describe the underlying open-source Tournament Manager software that runs them.

Game companies avoid sharing their game data with external researchers. Only a few research groups have been granted limited access to game data so far. The reluctance of these companies to make data publicly available limits the wide use and development of data mining techniques and artificial intelligence research specific to the game industry. In this paper, we developed and implemented an international competition on game data mining using commercial game log data from one of the major game companies in South Korea: NCSOFT. Our approach enabled researchers to develop and apply state-of-the-art data mining techniques to game log data by making the data open. For the competition, data were collected from Blade & Soul, an action role-playing game, from NCSOFT. The data comprised approximately 100 GB of game logs from 10 000 players. The main aim of the competition was to predict whether a player would churn and when the player would churn during two periods between which the business model was changed to a free-to-play model from a monthly subscription. The results of the competition revealed that highly ranked competitors used deep learning, tree boosting, and linear regression.

In general, the analysis of microarray data requires two steps: feature selection and classification. From a variety of feature selection methods and classifiers, it is difficult to find optimal ensembles composed of any feature-classifier pairs. This paper proposes a novel method based on the evolutionary algorithm (EA) to form sophisticated ensembles of features and classifiers that can be used to obtain high classification performance. In spite of the exponential number of possible ensembles of individual feature-classifier pairs, an EA can produce the best ensemble in a reasonable amount of time. The chromosome is encoded with real values to decide the weight for each feature-classifier pair in an ensemble. Experimental results with two well-known microarray datasets in terms of time and classification rate indicate that the proposed method produces ensembles that are superior to individual classifiers, as well as other ensembles optimized by random and greedy strategies.

The evolutionary approach for gaming is different from the traditional one that exploits knowledge of the opening, middle, and endgame stages. It is therefore sometimes inefficient to evolve simple heuristics that may be created easily by humans because it is based purely on a bottom-up style of construction. Incorporating domain knowledge into evolutionary computation can improve the performance of evolved strategies and accelerate the speed of evolution by reducing the search space. In this paper, we develop an evolutionary Othello player with the systematic insertion of opening knowledge into the framework of evolution. The probability of opening selection is coming from the expert’s opening list. Preliminary experimental results show that the proposed method is promising for generating better strategies for Othello players.

Bayesian network (BN) is a useful tool to represent joint probability distribution in the form of graphical model providing flexible inference and uncertainty handling. If there is enough knowledge about domain, it is possible to design the structure and parameters of BN by expert. Also, it can be learned from massive dataset with statistical learning algorithm. Usually, because the search space of Bayesian networks is relatively huge compared to the other models, evolutionary algorithms have been used to find optimal structure and parameters by many researchers. In this paper, we have focused on the topic of adaptation of constructed models for better performance. If there are a number of models constructed or learned by different experts or sources, it is better to fuse them into one model by considering all the information of each model. However, the complexity of the integrated model is relatively higher than previous isolated models. Minimizing the complexity of the integrated model using evolutionary algorithm is proposed. After integrating models into single one, it needs to adapt to the new data from the environment. It is likely to provide wrong results to the newly generated data from the environment and slightly modifying the joint probability distribution is necessary. The refinement process is also guided by the evolutionary algorithm because the space of search is large. Experimental results on a benchmark network show that the proposed adaptation methods with evolutionary algorithm can perform better than heuristics or greedy approaches.

We review the applications of artificial life (ALife), the creation of synthetic life on computers to study, simulate, and understand living systems. The definition and features of ALife are shown by application studies. ALife application fields treated include robot control, robot manufacturing, practical robots, computer graphics, natural phenomenon modeling, entertainment, games, music, economics, Internet, information processing, industrial design, simulation software, electronics, security, data mining, and telecommunications. In order to show the status of ALife application research, this review primarily features a survey of about 180 ALife application articles rather than a selected representation of a few articles. Evolutionary computation is the most popular method for designing such applications, but recently swarm intelligence, artificial immune network, and agent-based modeling have also produced results. Applications were initially restricted to the robotics and computer graphics, but presently, many different applications in engineering areas are of interest.

The evolutionary approach for gaming is different from the traditional one that exploits knowledge of the opening, middle, and endgame stages. It is, therefore, sometimes inefficient to evolve simple heuristics that may be created easily by humans because it is based purely on a bottom-up style of construction. Incorporating domain knowledge into evolutionary computation can improve the performance of evolved strategies and accelerate the speed of evolution by reducing the search space. In this paper, we propose the systematic insertion of opening knowledge and an endgame database into the framework of evolutionary checkers. Also, the common knowledge that the combination of diverse strategies is better than a single best one is included in the middle stage and is implemented using crowding algorithm and a strategy combination scheme. Experimental results show that the proposed method is promising for generating better strategies.

DNA microarray technology provides a format for the simultaneous measurement of the expression level of thousands of genes in a single hybridization assay. One exciting result of microarray technology has been the demonstration that patterns of gene expression can distinguish between tumors of different anatomical origins. Standard statistical methodologies in classification and prediction do not work well or even at all when N (the number of samples) < p (genes). Modification of existing statistical methodologies or development of new methodologies are needed for the analysis of cancer. Recently, designing artificial neural networks (ANNs) by evolutionary algorithms has emerged as a preferred alternative to the common practice of selecting the apparent best network. We propose an evolutionary neural network that classifies gene expression profiles into normal or colon cancer cell. Colon cancer is the second only to lung cancer as a cause of cancer-related mortality in Western countries. Colon cancer is a genetic disease, propagated by the acquisition of somatic alterations that influence gene expression. Experimental results on colon microarray data with evolutionary neural network show that the proposed method can perform better than other classifiers. Contribution of this article is applying evolutionary neural network to gene expression classification problem.

Recently, many researchers attempt to develop an effective SOM-based pattern recognizer for high performance classification. Structure adaptive self-organizing map (SASOM) is a variant of SOM that is useful to pattern recognition and visualization. Fusion of classifiers can overcome the limitation of a single classifier by complementing each other. Fuzzy integral is a combination scheme that uses subjectively defined relevance of classifiers. In this paper, fusion of SASOM's using fuzzy integral is proposed for Web mining problem. User profile represents different aspects of user's characteristics and needs an ensemble of classifier that estimate user's preference using Web content labeled by user as "like" or "dislike." The proposed method estimates the user profile using subsets of important features extracted from user-rated Web documents. Using UCI Syskill & Webert data, the method is tested and compared with other classifier including ID3, BP and naive Bayes classifier. Experimental results show that the fusion of SASOM's using fuzzy integral can perform better than not only previous studies but also majority voting of SASOM's.

Conventional evolutionary algorithms have a property that only one solution often dominates and it is sometimes useful to find diverse solutions and combine them because there might be many different solutions to one problem in real-world problems. Recently, developing checkers players using evolutionary algorithms has been widely exploited to show the power of evolution for machine learning. In this paper, we propose an evolutionary checkers player that is developed by a speciation technique called the "crowding algorithm". In many experiments, our checkers player with an ensemble structure showed better performance than non-speciated checkers players. A neural network is used to validate the game board, and a min-max search finds the optimal board. The neural network evaluator is evolved using the evolutionary algorithm.

In behavior-based robotics the control of a robot is shared between a set of purposive perception-action units, called behaviors. A major issue in the design of behavior-based control systems is the formulation of effective mechanisms for coordination of the behaviors' activities into strategies for rational and coherent behavior. There has been extensive work to construct an optimal controller for a mobile robot by evolutionary approaches such as genetic algorithm, genetic programming, and so on. In this line of research, we have also presented a method of applying CAM-Brain, evolved neural networks based on cellular automata (CA), to control a mobile robot. However, this approach has limitations to make the robot to perform appropriate behavior in complex environments. The multi module coordination method can make complex and general behaviors by combining several modules evolved or programmed, to do a simple behavior. In this paper, we coordinate several modules evolved to do a simple behavior by Maes's action selection mechanism. Maes (1989) has proposed a mechanism for action selection, which is reviewed here and is evaluated using a simulation environment. Experimental results show that this approach has potential to develop a sophisticated evolutionary neural controller for complex environments.