Neural Bayesian goal inference for symbolic planning domains

Abstract

There are several reasons for which one may aim to infer the short- and long-term goals of agents in diverse physical domains. As increasingly powerful autonomous systems come into development, it is conceivable that they may eventually need to accurately infer the goals of humans. There are also more immediate reasons for which this sort of inference may be desirable, such as in the use case of intelligent personal assistants. This thesis introduces a neural Bayesian approach to goal inference in multiple symbolic planning domains and compares the results of this approach to the results of a recently developed Monte Carlo Bayesian inference method known as Sequential Inverse Plan Search (SIPS). SIPS is based on sequential Monte Carlo inference for Bayesian inversion of probabilistic plan search in Planning Domain Definition Language (PDDL) domains. In addition to the neural architectures, the thesis also introduces approaches for converting PDDL predicate state representations to numerical arrays and vectors suitable for input to the neural networks. The experimental results presented indicate that for the domains investigated, in cases where the training set is representative of the test set, the neural approach provides similar accuracy results to SIPS in the later portions of the observation sequences with a far shorter amortized time cost. However, in earlier timesteps of those observation sequences and in cases where the training 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 method where SIPS uses a neural proposal based on the neural networks designed in this thesis could have the potential to combine the advantages of both goal inference approaches by improving the speed of SIPS inference while maintaining generalizability and high accuracy throughout the timesteps of the observation sequences.