Using Vector Representation of Propositions and Actions for STRIPS Action Model Learning
Using Vector Representation of Propositions and Actions for STRIPS Action Model Learning

Action model learning has become a hot topic in knowledge engineering for automated planning. A key problem for learning action models is to analyze state changes before and after action executions from observed "plan traces". To support such an analysis, a new approach is proposed to partition propositions of plan traces into states. First, vector representations of propositions and actions are obtained by training a neural network called Skip-Gram borrowed from the area of natural language processing (NLP). Then, a type of semantic distance among propositions and actions is defined based on their similarity measures in the vector space. Finally, k-means and k-nearest neighbor (kNN) algorithms are exploited to map propositions to states. This approach is called state partition by word vector (SPWV), which is implemented on top of a recent action model learning framework by Rao et al. Experimental results on the benchmark domains show that SPWV leads to a lower error rate of the learnt action model, compared to the probability based approach for state partition that was developed by Rao et al.
Action model learning has become a hot topic in knowledge engineering for automated planning. A key problem for learning action models is to analyze state changes before and after action executions from observed "plan traces". To support such an analysis, a new approach is proposed to partition propositions of plan traces into states. First, vector representations of propositions and actions are obtained by training a neural network called Skip-Gram borrowed from the area of natural language processing (NLP). Then, a type of semantic distance among propositions and actions is defined based on their similarity measures in the vector space. Finally, k-means and k-nearest neighbor (kNN) algorithms are exploited to map propositions to states. This approach is called state partition by word vector (SPWV), which is implemented on top of a recent action model learning framework by Rao et al. Experimental results on the benchmark domains show that SPWV leads to a lower error rate of the learnt action model, compared to the probability based approach for state partition that was developed by Rao et al.