Solving Interactive POMDPS in Julia
thesisposted on 06.08.2019 by Iacopo Olivo
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
Acting optimally in partially observable multi-agent stochastic domains is a growing topic in the Artificial Intelligence community and several solutions have been proposed. InteractivePOMDPs are one of the most complete solutions but it is highly susceptible to course of dimensionality and course of history. Several teams are proposing algorithms to overcome these difficulties. In this work it is proposed the framework Julia.POMDPs in order to standardize the development and testing of such solving algorithms (Chapter 3). The framework introduces ways to declare I-POMDP agents and how to define an agent hierarchy. Julia.IPOMDPs takes advantage of Julia.POMDPs to provide solutions to POMDPs. This project also proposes a new way to solve I-POMDPs by reducing them to POMDPs (Chapter 4) and solving them by means of Julia.POMDPs. However, the solver could not provide the expected precision due to loss of information in the conversion. The on-line solver is tested by the redefinition of the multi-agent tiger game. Tests and results are analyzed in Chapter 5. The tests are used in order to show the simplicity of defining several different frames and problem setups.