Latent Dynamics Model Learning as Multi-objective Optimization with Augmented Tchebyshev Function

Abstract

With the recent development of deep learning technology, model learning of low-dimensional potential dynamics embedded in high-dimensional observation data has attracted much attention for complicated robotic control on unknown environment. In previous research, this model learning is formulated as a variational lower bound maximization problem of the marginalized log-likelihood of observed time-series data. From the viewpoint of multi-objective optimization, this problem can be interpreted to be a linearly weighted sum of multiple objective functions, which can only find the convex part of the Pareto front even by optimizing the weight parameters. To find all the Pareto front regardless of its shape, this paper proposes a new model learning framework based on multi-objective optimization with the augmented Tchebyshev scalarization. In the simulation of the object manipulation problem by a robot manipulator with camera images as observation, the effectiveness of the proposed method is demonstrated through Baysian optimization, which efficiently explores one preferred solution on the Pareto front.