Basic Concepts: Key terms for understanding our research

To help you understand our research, we’ll explain four important terms that frequently appear.

---

Conventional dialogue devices like standard smart speakers typically respond only when spoken to by humans. In contrast, our goal is to develop an Intelligent Dialogue Agent (IDA) that proactively initiates warm conversations, acting as an empathetic companion similar to a family member or friend.

Using Reinforcement Learning (RL), the dialogue agent learns individual user preferences and interests (e.g., sports, politics, entertainment). Furthermore, by integrating state-of-the-art natural language processing techniques such as Sentiment Analysis and Distributed Text Representation, it comprehends the sentiment (positive/negative) of news items and delivers fresh, engaging topics in a timely manner without repeating past conversations.

Additionally, this dialogue agent serves as the core unit of a Comprehensive Preventive Care System (CPCS) linked with in-room motion sensors and cognitive training routines. While unobtrusively monitoring daily living activities, it prevents social isolation among older adults through engaging conversations, creating synergistic effects that boost motivation for physical exercise and brain training.

---

Preventing dementia and motor function decline (such as fall risks) requires daily moderate physical activity and cognitive stimulation. However, sustaining repetitive physical exercises or simple math drills over long periods is challenging. To address this, we develop integrated exergaming technologies that fuse exercise and gaming into a single seamless experience.

For instance, in a "Red Light, Green Light" (Daruma-san ga Koronda in Japanese) game played with a communication robot, users control an on-screen avatar by stepping while seated in a chair. In our latest Cogni Puzzle, users perform body gestures (e.g., spreading arms sideways, clapping overhead) in front of a built-in camera to rotate and clear falling puzzle blocks.

Here, the AI real-time evaluates not only game proficiency but also physical exertion via the Borg Scale (Rating of Perceived Exertion) and movement form accuracy (e.g., knee height). Powered by Reinforcement Learning (RL), the AI automatically adjusts game difficulty and exercise load, preventing users from quitting due to excessive difficulty or losing interest from lack of challenge. By embedding gamification concepts, we enable sustainable, enjoyable health habits.

---

Modern society faces serious health issues, including declining mental and physical health among students and workers due to academic/work performance strain or overwork, as well as an increasing number of older adults requiring long-term care due to irregular routines, falls, or dementia. We conduct research on a Mental-Physical States Estimation System designed to detect subtle health changes at the pre-symptomatic ("Mibyou") stage and prevent health deterioration before symptoms fully emerge.

This system collects and integrates biological signals—such as R-R Interval (RRI) and sleep patterns from wearable devices like smartwatches—alongside daily activity logs (lifelogs) and environmental data like ambient temperature and humidity.

Using probabilistic models and advanced AI techniques on this gathered big data, the system analyzes complex causal dependencies, such as identifying that a combination of sleep deprivation and specific weather conditions causes a sharp increase in physical/mental burden two to three days later. Upon predicting near-future health risks, the AI delivers personalized recommendations via smartphones, such as "Consider getting extra rest today" or "Hydration and light stretching will be effective." Our long-term vision is a future where individuals are naturally safeguarded simply by wearing smart devices, minimizing manual user data entry.

---

Real-world phenomena do not always follow simple deterministic rules like "if X occurs, Y will always happen." For AI to make intelligent and safe decisions in human society, it requires probabilistic reasoning capabilities under uncertainty. Our laboratory researches the underlying theory and algorithm acceleration/optimization of Bayesian Networks (BNs), which serve as the mathematical foundation for AI intelligence.

As data grows larger and more complex, massive computational demands can create processing bottlenecks in AI. To solve this, we developed Stepwise Structure Learning with Probabilistic Pruning (SSL-Pro), which eliminates redundant computations based on data dependencies to achieve fast network structure learning.

Furthermore, we have achieved advanced theoretical contributions, including guaranteed-accuracy probabilistic inference algorithms (Extended-LBPC: Ex-LBPC) that mathematically verify the reliability of AI predictions, as well as Mixture of Bayesian Networks for rapid robot adaptation to changing environments. These mathematical foundations reliably support the robust performance of our preventive care AI and health prediction systems behind the scenes.

---