curiosity-driven

Compression progress measures how much an agent's world model improves at predicting observations. The agent receives intrinsic reward equal to the reduction in description length, encouraging exploration of patterns that improve the model rather than random novelty. This satisfies curiosity-driven learning by directly measuring learnable structure.

Yes. The Skill integrates with Python frameworks PyTorch and JAX to compute intrinsic rewards. You pair a differentiable world model and compressor, then calculate rewards as description-length improvement to guide action selection in model-based control.

Novelty-based curiosity rewards the agent for visiting unfamiliar states. Compression-based curiosity rewards improvement in the world model's ability to compress observations, focusing exploration on learnable patterns that reduce description length rather than mere surprise.

Create a CuriosityDrivenAgent with a differentiable world model and compressor. Define a description-length-based reward where reward equals the previous description length minus the current length. Integrate with PyTorch or JAX to compute intrinsic rewards and select actions that maximize compression progress.

You need observations from the environment and a differentiable world model capable of predicting future states. The compressor must operate on model outputs to measure description length. Both components must support gradient computation in PyTorch or JAX.

Use compression progress when you need intrinsic motivation that targets learnable structure in the environment. It suits model-based reinforcement learning and control scenarios where the agent benefits from discovering patterns that improve world-model accuracy.