Weltmodell

Weltmodell, or world model in English, is a term used in cognitive science and artificial intelligence to denote an internal representation of the external world that an agent uses to predict the consequences of actions and to plan behavior. In cognitive science, world models are part of predictive processing and forward models that enable perception, motor control, and navigation by simulating likely sensory outcomes of actions. They are learned from experience, can be probabilistic, multi-modal, and organized hierarchically.

In AI and reinforcement learning, a world model is a model of the environment that supports planning and imagined trajectories. A typical world model includes a dynamics model p(s'|s,a) predicting future states, an observation model p(o|s) for sensor readings, and a reward model r(s,a). Once learned, this model allows a controller to simulate outcomes and optimize policies with fewer real-world interactions, improving sample efficiency.

Notable work includes the World Models framework (Ha and Schmidhuber, 2018), which uses a variational autoencoder to compress observations into a latent state, a recurrent network to predict future latent states and rewards, and a controller that operates in the latent space. The approach demonstrated planning and control in simulated tasks and influenced subsequent research in model-based reinforcement learning and generative world representations.

Definitions and emphasis vary by field, but the core idea remains the same: a Weltmodell is an internal or explicitly learned model that enables anticipation, planning, and action by simulating how the world responds to choices. Limitations include model bias and challenges in scaling to complex environments.