Learn Unity ML-Agents - Fundamentals of Unity Machine Learning

Fixing the observation of state

Making sure you are capturing the relevant observation of state for your agent is critical to successfully train an agent. In most of the earlier examples, the way in which we built the observation of state was quite simplistic, but as you can now appreciate, an agent's state can be quite substantial. In fact, some RL problems currently being tackled have states exceeding the number of atoms in the known universe—yes, you read that right. We broached this subject in the last chapter, where we demonstrated how the agent observations could be mapped as inputs onto a neural network. When setting up the Unity external brain trainer, it will be essential that you understand the how or what of things that an agent ...