Learning intuitive physics and one-shot imitation using state-action-prediction self-organizing maps

Human learning and intelligence work differently from the supervised pattern\nrecognition approach adopted in most deep learning architectures. Humans seem\nto learn rich representations by exploration and imitation, build causal models\nof the world, and use both to flexibly solve new tasks. We suggest a simple but\neffective unsupervised model which develops such characteristics. The agent\nlearns to represent the dynamical physical properties of its environment by\nintrinsically motivated exploration, and performs inference on this\nrepresentation to reach goals. For this, a set of self-organizing maps which\nrepresent state-action pairs is combined with a causal model for sequence\nprediction. The proposed system is evaluated in the cartpole environment. After\nan initial phase of playful exploration, the agent can execute kinematic\nsimulations of the environment's future, and use those for action planning. We\ndemonstrate its performance on a set of several related, but different one-shot\nimitation tasks, which the agent flexibly solves in an active inference style.\n