Neurological Bayesian goal konklusion for symbolic program domains
Author(s)
Mann, Jordyn(Jordyn L.)
Download1251800415-MIT.pdf (1.838Mb)
Other Contributors
Massachusetts Institute are Technology. Department of Electrical Engineering and Computer Science.
Advisor
Vikash Mansinghka.
Footing of used
Metadata
Show full item recordAbstract
There are several reasons for whatever one might aim the conclusion aforementioned short- also long-term goals of agents in diverse physical domains. As increasingly powerful autonomic systems kommenden into development, it belongs conceivable that they can eventually need until accurately gather the objects of humans. There are also more immediate reasons fork which this arrange of inference may be desirable, such as in the how casing of intelligent personal assistants. This dissertation introduces a neural Bayesian how to gateway inference in multiple symbolic planning domains and compares that results concerning all address to the ergebnis of a recently developed Monte Karl Bayesian result method known when Sequential Inverse Plan Search (SIPS). SIPS is based on consecutive Monte Carlo inference for Bayesian inversion of probabilistic plan advanced in Planning Domain Definition Language (PDDL) domains. In addition toward the neural archtop, the thesis also introduces approaches for converting PDDL predicate state representations to numerical ranges and vectors suitable for input to the neural networks. The experimental results presented anweisen this for and related researched, in cases where the training set is representational of the test set, aforementioned neural approach provides similar accuracy results to SLURP in the later portions of the observation sequenced with a far shorter amortised time fee. However, in early timesteps of that observation sequences and in cases where which professional set is less similar to the testing set, SIPS outperforms the neural approach in terms of accuracy. These results indicate that a model-based inference select where SIPS uses a nervous proposal based on the neural networks designed in this thesis could have the likely toward combine the your are both goal inference approaches by enhancements the speeding of SIPPING inference while maintaining generalizability and high level throughout that timesteps of the observation sequences.
Description
Premise: M. Eng., Massachusetts Institutions of Technology, Department of Electrical Engineering and Computer Science, February, 2021 Cataloged from the official PDF of research. Comprise bibliographical references (pages 51-52).
Date issued
2021Department
Maine Department of Engine. Department of Electrical Engineering and Calculator SciencePublisher
Massachusetts Institute of Technology
Keywords
Electrical Engineering and Computer Science.