Young researcher's career development project

Project Code: DOK-NPOO-2023-10-1790, Croatian Science Foundation

The main goal of this project is to guide a young doctoral researcher toward the successful defense of its doctoral dissertation. The additional goal is to develop a computational model that enables a software agent to learn during interaction by using nonverbal communication cues and by analyzing the available information space. The reasoning process of the software agent will be based on the analysis of interaction and the information space in which the interaction is embedded, as well as on continuous adaptation to changes characteristic of real-world environments.

Human–robot interaction represents a rapidly evolving scientific field. New methods based on deep learning enable systems to utilize stored knowledge during decision-making processes. Current decision hypotheses may be confirmed (strengthened) or rejected (weakened) through interaction. Therefore, interaction itself can represent an important mechanism through which a system adapts to changes.

The flow of interaction thus provides information that the computational model can use when forming or rejecting decisions regarding subsequent interaction steps. Traditional machine learning techniques mainly rely on pre-prepared datasets, where the learning process is completed in advance and remains unchanged while the system is in use (Batch Learning). As long as environmental conditions remain at least approximately the same, the model should be capable of producing satisfactory results.

However, a defining characteristic of the real world is continuous change, meaning that variability is constant. Consequently, the prediction accuracy of the model decreases as differences from the original training conditions increase. Adaptive AI incorporates new data from the working environment during interaction in order to generate more accurate insights into the state of the environment. Newly collected and utilized data enable adaptations and updates of the underlying code, allowing the AI system to dynamically retrain, learn, and improve while tracking environmental changes.

Carefully collected data may reflect changes in the behavior of the person involved in the interaction, meaning that the AI system can continuously adapt its interaction strategy while the interaction is taking place. Such an approach is necessary for timely and appropriate responses, since the environment is dynamic and the perspective of the interacting person itself represents a source of uncertainty due to its subjective nature.

The new computational model based on adaptive learning will therefore enable the creation and maintenance of mutual understanding through interaction by using the information space as a shared pool of continuously updated information. AI methods based on real-time learning are considered the next evolutionary phase in the development of computational models.