RT Dissertation/Thesis
T1 Learning of Geometric-based Probabilistic Self-Awareness Model for Autonomous Agents
A1 Iqbal, Hafsa
AB In the emerging domain of self-aware and autonomous systems, the causal representation ofthe variables and the learning of their dynamics to make inferences of the future state at multipleabstraction levels in successive temporal slices, are getting attention of the researchers.This work presents a novel data-driven approach to learn the causal representation of thedynamic probabilistic graphical model. The proposed method employed the geometric-basedapproach to define the set of clusters of similar Generalized State (GS) space as linearattractors. Clustering of data variables corresponding to the linear attractors defines a set ofswitching vocabulary, which provides the higher-level representation of the graphical model,i.e., discrete and continuous levels. The transitions between the switching vocabulary arerepresented with the transition matrix, estimated from the temporal data series in switchingmodels based on GSs. Each learned representation of the dynamic probabilistic graphicalmodel is stored in Autobiographical Memory (AM) layers. A Markov Jump Particle Filter(MJPF) is proposed to make inferences at multiple abstraction levels of graphs whichfacilitates the detection of anomalies. Anomalies indicate that the agent encounters newexperiences which can be learned incrementally and evolve new layers of AM.The proposed approach is extended for the learning of interactions between autonomousagents to make them self-aware. In Low dimensional case, data from the odometry trajectoriesand the control parameters, i.e., steering angle and rotors’ velocity, is employed. However,data from the LiDAR, i.e., 3D point clouds, is used for the high-dimensional case. The deeplearning approach, such as 3D Convolutional Encode-Decoder together with the transferlearning employed to extract the features from the LiDAR’s point clouds. A similar learningapproach (mentioned above) is employed to detect anomalous situations. Three predictivemodels, i.e., piecewise nonlinear, piecewise linear, and nonlinear models, are proposed toanalyse the multiple abstraction level anomalies, i.e., continuous level, discrete level, andvoxel level. Concurrently, the public KITTI dataset from the complex/urban environment isemployed to validate the proposed methodology. Qualitative and quantitative analysis of theproposed methodology is perform by employing the anomaly measurements and the ROCcurves to estimate the accuracy, respectively.
YR 2022
FD 2022-04
LK https://hdl.handle.net/10016/35943
UL https://hdl.handle.net/10016/35943
LA eng
NO Mención Internacional en el título de doctor
DS e-Archivo
RD 27 jul. 2024