Academia.eduAcademia.edu

Figure 13 – uploaded by Ramon Pereira

See full PDFdownloadDownload figure
Figure 4.4: Cost Difference approach example. vhile J~ computes trajectories that aim to avoid achieving the observed states in Obs.

Figure 4 4: Cost Difference approach example. vhile J~ computes trajectories that aim to avoid achieving the observed states in Obs.

Related Figures (28)

Figure 2.1: Ordered Landmarks for a BLOCKS-WORLD problem instance. Connected boxes represent conjunctive landmarks. set of facts that must be true together at some point in every valid plan to achieve a goal.
Figure 2.1: Ordered Landmarks for a BLOCKS-WORLD problem instance. Connected boxes represent conjunctive landmarks. set of facts that must be true together at some point in every valid plan to achieve a goal.
Figure 2.2: Plan Recognition as a superset of Goal Recognition [63].
Figure 2.2: Plan Recognition as a superset of Goal Recognition [63].
Figure 2.3: Goal Recognition example for BLOCKS-WORLD domain.
Figure 2.3: Goal Recognition example for BLOCKS-WORLD domain.
observer and the observed agent. Figure 3.1: Problem overview to goal recognition in incomplete domain models. from previous planning approaches |96, 60]. Figure 3.1 shows an overview of how we define the By combining the various notions of planning in incomplete domains [96, 60] and ob-
observer and the observed agent. Figure 3.1: Problem overview to goal recognition in incomplete domain models. from previous planning approaches |96, 60]. Figure 3.1 shows an overview of how we define the By combining the various notions of planning in incomplete domains [96, 60] and ob-
Figure 3.2: ORPG for the abstract incomplete planning problem presented in Example 2.1. Green arrows represent preconditions, Orange arrows represent add effects, and Purple dashed arrows represent possible add effects. Light-Blue boxes represent the set of definite landmarks and Light yellow boxes represent the set of possible landmarks. Grey hexagons represent actions. Goal Recognition Heuristics over Incomplete Domain Models We now develop novel goal recognition heuristics that rely on landmarks over incom.
Figure 3.2: ORPG for the abstract incomplete planning problem presented in Example 2.1. Green arrows represent preconditions, Orange arrows represent add effects, and Purple dashed arrows represent possible add effects. Light-Blue boxes represent the set of definite landmarks and Light yellow boxes represent the set of possible landmarks. Grey hexagons represent actions. Goal Recognition Heuristics over Incomplete Domain Models We now develop novel goal recognition heuristics that rely on landmarks over incom.
Figure 3.3: Heuristic Goal Recognition over Incomplete Domain Models.
Figure 3.3: Heuristic Goal Recognition over Incomplete Domain Models.
Table 3.1: The average number of possible complete domain models |((D))| for all domains we use in our experiments.
Table 3.1: The average number of possible complete domain models |((D))| for all domains we use in our experiments.
Table 3.2: Experimental results for our ablation study, comparing the baseline approaches hy. and Muniq (APPENDIX A) against our enhanced heuristics using various combinations of landmark types. This table aggregates the average results for all metrics over all datasets for all levels of observability.
Table 3.2: Experimental results for our ablation study, comparing the baseline approaches hy. and Muniq (APPENDIX A) against our enhanced heuristics using various combinations of landmark types. This table aggregates the average results for all metrics over all datasets for all levels of observability.
yaseline approaches (APPENDIX A), in all variations of domain incompleteness and obsery-
yaseline approaches (APPENDIX A), in all variations of domain incompleteness and obsery-
Figure 3.5: Fl-score average of all evaluated approaches over the datasets varying the domain incompleteness. We now present our second set of experiments, comparing the results of our enhanced
Figure 3.5: Fl-score average of all evaluated approaches over the datasets varying the domain incompleteness. We now present our second set of experiments, comparing the results of our enhanced
Figure 3.6: ROC space analysis for all four percentage of domain incompleteness, comparing our enhanced heuristic approaches against the baselines hy. and hunig (APPENDIX A).
Figure 3.6: ROC space analysis for all four percentage of domain incompleteness, comparing our enhanced heuristic approaches against the baselines hy. and hunig (APPENDIX A).
Figure 4.1: Problem overview to goal recognition over nominal models. Artificial Neural Networks have shown to be very effective at learning and approxi-
Figure 4.1: Problem overview to goal recognition over nominal models. Artificial Neural Networks have shown to be very effective at learning and approxi-
Figure 4.2: An example of a neural network with 4 layers (input layer, 2 hidden layers, and output layer), adapted from [85]. Arrows represent full connections between the input and output nodes. Every hidden layer output unit is passed through a ReLU unit. Dashed arrows are optional dense connections. architecture reported by Say et al. [85], in which we use to represent nominal models. her advantage that, as demonstrated in [98, 85], DNNs can be directly used in Equation 2.6,
Figure 4.2: An example of a neural network with 4 layers (input layer, 2 hidden layers, and output layer), adapted from [85]. Arrows represent full connections between the input and output nodes. Every hidden layer output unit is passed through a ReLU unit. Dashed arrows are optional dense connections. architecture reported by Say et al. [85], in which we use to represent nominal models. her advantage that, as demonstrated in [98, 85], DNNs can be directly used in Equation 2.6,
Figure 4.3: Nominal Mirroring approach example. To analyze the computational complexity of 7 MIRRORING, we use as a baseline Vered
Figure 4.3: Nominal Mirroring approach example. To analyze the computational complexity of 7 MIRRORING, we use as a baseline Vered
Figure 4.5: An example of a goal recognition problem for the 2D-LQR-—Navigation domain with a single vehicle. function to compute the trajectories for the possible goals considering the observations. Like in Section 4.2, we learn the system dynamics of models (transition function)
Figure 4.5: An example of a goal recognition problem for the 2D-LQR-—Navigation domain with a single vehicle. function to compute the trajectories for the possible goals considering the observations. Like in Section 4.2, we learn the system dynamics of models (transition function)
Linear LQR-Based Domains Table 4.2: Experimental results of our recognition approaches over both actual and nominal models for the LQR-based domains. M represents the model type (A represents actual models, and N represents nominal models), % Obs is observation level, and N is the total number of observed states. Note that the average number of goal hypothesis |G| in the datasets is 5, and the planning horizon H for these datasets is 100.
Linear LQR-Based Domains Table 4.2: Experimental results of our recognition approaches over both actual and nominal models for the LQR-based domains. M represents the model type (A represents actual models, and N represents nominal models), % Obs is observation level, and N is the total number of observed states. Note that the average number of goal hypothesis |G| in the datasets is 5, and the planning horizon H for these datasets is 100.
Non-Linear Navigation Domains Table 4.3: Experimental results of our recognition approaches over both actual and nominal models for the non-linear navigation domains (2D-NAV and 3D-NAV). M represents the model type (A represents actual models, and N represents nominal models), % Obs is observation level, and N is the total number of observed states. Note that the average number of goal hypothesis |G| in these datasets is 4, and the planning horizon H for these datasets is 20. With respect to recognition time, the average time per goal recognition problem for
Non-Linear Navigation Domains Table 4.3: Experimental results of our recognition approaches over both actual and nominal models for the non-linear navigation domains (2D-NAV and 3D-NAV). M represents the model type (A represents actual models, and N represents nominal models), % Obs is observation level, and N is the total number of observed states. Note that the average number of goal hypothesis |G| in these datasets is 4, and the planning horizon H for these datasets is 20. With respect to recognition time, the average time per goal recognition problem for
Figure APP] FINDIX A.1: Ordered fact landmarks extracted for the word RED. Fact landmarks that must be true together are represented by connected boxes. Connected boxes in grey represent ac hieved fact landmarks. ] Edges represent prerequisites between landmarks. achieved landmarks for the word RED in gray.
Figure APP] FINDIX A.1: Ordered fact landmarks extracted for the word RED. Fact landmarks that must be true together are represented by connected boxes. Connected boxes in grey represent ac hieved fact landmarks. ] Edges represent prerequisites between landmarks. achieved landmarks for the word RED in gray.
ENDIX A.1: Listing APP] Extracted fact landmarks for the |] Figure 2.3 and their respective uniqueness value. BLOCKS-WORLD example in
ENDIX A.1: Listing APP] Extracted fact landmarks for the |] Figure 2.3 and their respective uniqueness value. BLOCKS-WORLD example in
Table APPENDIX B.4: Performance results for ablation study, comparing the baseline approaches against our heuristics using some possib combinations of landmark types, separated by observability (70% and 100%).
Table APPENDIX B.4: Performance results for ablation study, comparing the baseline approaches against our heuristics using some possib combinations of landmark types, separated by observability (70% and 100%).
2D-LQR-Navigation Single Vehicle (Linear Domain) Table APPENDIX § C.2: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 2D—-LQR- Navigation single vehicle linear domain.
2D-LQR-Navigation Single Vehicle (Linear Domain) Table APPENDIX § C.2: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 2D—-LQR- Navigation single vehicle linear domain.
2D-LQR-Navigation with Multiple Vehicles (Linear Domain) Table APPENDIX C.3: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 2D—-LQR- Navigation linear domain with multiple vehicles.
2D-LQR-Navigation with Multiple Vehicles (Linear Domain) Table APPENDIX C.3: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 2D—-LQR- Navigation linear domain with multiple vehicles.
Table APPENDIX C.4: Experimental results of our recognition approaches over both actual and nominal models for the 2D-NAV non-linear domain. for linear domains, especially for offline goal recognition.
Table APPENDIX C.4: Experimental results of our recognition approaches over both actual and nominal models for the 2D-NAV non-linear domain. for linear domains, especially for offline goal recognition.
Table APPENDIX C.5: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 3D-NAV non-linear domain.
Table APPENDIX C.5: Experimental results of our recognition approaches nMIRRORING and A(Obs,G), over both actual and nominal models for the 3D-NAV non-linear domain.
Related topics:
Computer Science
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved