Automated Planning with Hybrid Domain Models: A Method to Improve Continuous Process Descriptions

Recent advances in Automated Planning not only involve the improvements in planning efficiency but also the enhancement in granularity in which planning domains are modelled. A significant progression is in the move from discrete domain models to mixed discrete-continuous models i.e. hybrid domain models. While planning with hybrid domains has been studied for decades, the knowledge engineering of those domain models is still a challenge, particularly for real-time complex domains. It is imperative to understand how to effectively and efficiently formulate the planning models to achieve maximum productivity with minimum wasted effort or cost. One of the main engineering challenges of hybrid domain models involves encoding the frequent fluctuation of underlying processes with continuous updates in the world state. The occurring numerical changes with the variation of parameters can be too complex to be formulated accurately by human manual efforts.

This thesis proposes a method utilising machine learning techniques which results in the formulation of a run-time representative estimation of continuous changes in varied process parameters. The method incorporates statistical analysis to acquire process models from real-world data for hybrid planning domains. We assume that domain knowledge has been already encoded in an initial hybrid domain model (in this thesis, that is a PDDL+ model). We then use the method to create an improved process model (within the same encoding language of PDDL+) which is embedded into the process specification of the original, pre-engineered domain model.

By exploiting the quantitative data from hybrid planning domains, firstly the proposed approach, with the help of statistical methods, identifies the associations (i.e. dependencies or inter-dependencies) between a single outcome and single/multiple predictor numeric variables in the underlying process. Based on the deduced statistical relationship among variables, the appropriate linear regression technique with corresponding statistical tests are nominated and implemented to formulate the process model. Then the constructed process model is embedded/adjusted into the pre-engineered hybrid domain models. The learned process model, in the form of a mathematical function, automatically approximates/adjusts the quantity of an outcome variable with the continuous variations in different predictor features in order to efficiently and accurately control the corresponding process in hybrid planning domains.

To empirically evaluate our approach, we utilise pre-engineered models (in PDDL+) of an Urban Traffic Control (UTC) domain and a Coffee domain. For the UTC domain, we experiment with the real-time traffic data that is collected from AIMSUN simulator utilised in the SimplifAI project (McCluskey, Vallati and Franco 2017). Besides, for coffee domain, we collect the real-time data from an observational study that is conducted by Easthope (2015). The evaluation results demonstrate that the automatically learned values of numeric process variables by our method are more rational than the formulation values of process variables declared statically in the original domain models. Besides, it reveals that the learned process models can provide more accurate simulation output, which can consequently lead to higher-quality plans. Along with that, by automatically identifying the effective process variables and removing the irrelevant ones from the learned process models, it can assist the knowledge engineering tasks of modelling/adjusting the dynamically changing process variables with their values in the hybrid planning domains, without declaring them statically.