Acquiring Information from Bayesian Surprise in Cognitive Linear Gaussian Dynamic Systems

The development of autonomous systems (AS) requires cognitive entities capable of acquiring information, learning, planning, and reasoning that eventually adapt to environmental uncertainties. Since surprising events encourage learning and information-seeking in biological agents, this research suggests surprise as the intrinsic motivator to design cognitive dynamic systems undergoing autonomous behavior. Amongst several definitions of surprise, this thesis proposes the Bayesian surprise to express information utility and to guide novelty acquisition from uncertain measurements by considering the impact of new data on changing prior beliefs. By adopting Gaussian noise-driven linear dynamic models to represent the environment, the Bayesian surprise drives the system to estimate future environmental states from noisy measurements and execute a decision to minimize the estimation error over time. This research implements a novel planning algorithm, where the system computes the contribution of prospective actions---which are available from a library---to state estimation and selects the one that maximizes the expectation of Bayesian surprise. Theoretical analysis shows that the action corresponding to the highest expectation of Bayesian surprise conveys the maximum information and reduces the state estimation error. Furthermore, a learning and planning algorithm is derived by intrinsically assigning novel rewards which are developed based on the expectation of Bayesian surprise. This thesis proposes reward functions inspired by credibility measures in estimation/control theory to evaluate the consequences of past actions. The performances of the proposed algorithms are compared to the state-of-the-art for numerous experiments. For a cognitive linear Gaussian dynamic application (i.e., cognitive radar), results show that the proposed planning algorithm outperforms its competitors with respect to the mean square relative error when one-step and multiple-step planning are considered. In comparison to alternative methods, the learning and planning algorithm implemented by the proposed surprise-based rewards significantly improves the state estimation performance. Also, results indicate that multiple-step planning does not necessarily lead to lower error, mainly when the environment changes abruptly.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.