Understanding brain agents and academy
Abhilash Majumder · Apress eBooks · 2020
The brain architecture is an important aspect of the ML Agents Toolkit. In the previous chapter, we installed the ML Agents Toolkit and also learned briefly about this architecture. Internally the ML Agents Toolkit uses three different kinds of brains, with the addition of a player brain that is controlled by the user. We concern ourselves with understanding the inner workings of certain scripts in the ML Agents package, which uses the neural networks trained in Tensorflow in Unity Agents. Since we have tried to get a glimpse of deep Q-learning as the only deep RL algorithm as of now, we can also use this algorithm to train the brain of the agent. Whereas in Unity ML Agents, the default algorithm for the internal brain is proximal policy operation (PPO), which is robust and has a comfortable balance between ease of implementation, sample tuning, and complexity, we will explore different algorithms that will be used as the brain for the agent. In this section, we will have a deep insight into the brain architecture and all the associated C# scripts associated with it, including the different aspects of model training and hyperparameter tuning. We will be building games using ML Agents in Unity.