Autonomous Learning of Commonsense Simulations

Parameter-driven simulations are an effective and efficient method for reasoning about a wide range of commonsense scenarios that can complement the use of logical formalizations. The advantage of simulation is its simplified knowledge elicitation process: rather than building complex logical formulae, simulations are constructed by simply selecting numerical values and graphical structures. In this paper, we propose the application of machine learning techniques to allow an embodied autonomous agent to automatically construct appropriate simulations from its real-world experience. The automation of learning can dramatically reduce the cost of knowledge elicitation, and therefore result in models of commonsense with breadth and depth not possible with traditional engineering of logical formalizations.