Modeling curiosity for intelligent agents
Date of Issue2014
School of Computer Engineering
In recent years, researchers have shown an increasing interest in studying various characteristics of intelligent agents, such as emotion, negotiation, trust, and persuasion. However, agents with these characteristics still lack the capability to direct their attention towards novelty and seek interestingness. In human beings, such capability allows us to achieve individual growth and a closer interpersonal relationship, which is governed by curiosity, the critical motivation associated with exploration, learning, and interest. Therefore, it can be envisioned that a computational model of curiosity may enable a learning agent to achieve better learning efficiency, a recommender agent to predict interesting items, or a pedagogical agent to demonstrate believable learning behaviors. These thoughts inspire us to model human-like curiosity for intelligent agents and study the impact brought by this novel characteristic. A computational framework of curiosity is developed in this thesis, which is largely inspired from the literature in psychology. Berlyne's and Wundt's theories on curiosity are chosen to be the basis for conducting this work due to their long lasting impact in the psychology of curiosity. Berlyne's theory identifies several key factors that govern the process of curiosity arousal, such as novelty, surprise, uncertainty, conflict, and complexity. Wundt's theory postulates an inverted U-shape relationship between the stimulation level and the hedonic value, which states that only optimal stimulation level results in curiosity. Based on these theories, we propose a computational framework of curiosity for intelligent agents. This framework is composed of a generic computational model of curiosity, a curiosity-driven learning algorithm, and a curiosity-driven recommendation algorithm. The computational model of curiosity consists of abstract functions and their interactions between each other. Representing computational models for intelligent agents with abstract functions makes them general enough to allow different implementations in different application contexts. The curiosity-driven learning algorithm realizes the computational model of curiosity in a fast neural learning agent: the extreme learning machine. Performance comparisons with other popular learning algorithms in the literature on benchmark classification problems demonstrate a superior learning and generalization ability of the curiosity-driven learning algorithm. The curiosity-driven recommendation algorithm realizes the computational model of curiosity in a social recommender agent. Experimental studies with large scale real world data sets show that curiosity significantly enhances the agent's recommendation diversity and coverage, while maintaining a sufficient level of accuracy. To evaluate the practical values of the proposed computational model of curiosity, we study it in the domain of virtual worlds for educational purposes; as it has been shown in the literature of education that curiosity is an important driving force for child development, scientific research, and educational achievements. Two types of pedagogical agents are chosen for study, a virtual peer learner and a virtual learning companion. The virtual peer learner is a Non-player Character that aims to provide a believable virtual environment without direct interactions with users, whereas the virtual learning companion directly interacts with users to enhance their learning experience. Studies on a curious peer learner through computer simulations show that curiosity enables a virtual peer learner to demonstrate human-like learning behaviors, which significantly enhances believability. Studies on a curious learning companion with human users show that curiosity of the virtual learning companion significantly improves the users' learning experience from several aspects, including learning gains, self-efficacy, and interest.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence