Why Train?
When you purchase a new NFT for the Gaming Competition, the neural network parameters making up its Core are randomly generated. This means that the neural network will initially perform random actions because the network has not yet developed any skills.
To prepare for battle, you must train the AI so it gains skills or learns an effective fighting policy.
Training is the process of changing the parameters in the neural network to make the AI act in a specific way - action - in a specific situation - state. The resulting combination of learned actions in different states is called a policy.
To help your AI master skills required to win, we provide a training environment called Imitation Learning.
To illustrate, imagine the only action you teach is punching. The AI will learn to punch no matter what situation it is in. Not a very effective strategy if you ask us...
An AI with a balanced policy is referred to as one that “generalizes” well. This is the ultimate objective of training.
Imitation Learning
Imagine you are a Sifu, Sensei or Coach, and the AI is your apprentice. You spar with your AI, and it learns to copy the moves you do in specific situations.
This is the premise of Imitation Learning (IL).
IL is an iterative loop 🔁 that has 4 steps:
🚰 Data Collection - This the start of the IL process, where you demonstrate actions to your AI. In this module, you are playing the game — not your NFT! You are actually creating a list of actions (in the observed states) for the AI to copy and learn from. In Machine Learning terms, you are creating the data set that the AI will train on. Remember, the AI can only train on data that you create for it, so the deliberate creation of useful data is very important and requires skill to achieve.
💻 Configuration - Once you have created the data, you can:
➡️ Define the training intensity.
- How much of the new data should be incorporated into your model?
- How much do you want the AI to remember from its previous training?
➡️ Pinpoint the features that you want the AI to focus its training on.
- For example, you can tell your AI to focus on relative positioning for this training session.
- Think about this as the coaching process after a sparring session - you apply learnings to specific situations with your fighter.
🏋️ Training - After collecting and configuring the data, you are able to train your AI to update its parameters. Your AI will evolve and adapt to a new policy.
🔍 Inspection - Your model is now updated, but do you actually know what it has learned and if it is getting better? The answer is yes! With the AI Inspector, you are able to see “inside the brain” of the AI. The Inspector provides a comprehensive view of what the AI’s action would be given a state, or, the current view of the policy mapping. Armed with this information, you can design and target specific improvement areas for the next iteration of training.
Simulation Mode
While training is a critical part of preparation, it is not a substitute for the dynamism of fighting against another AI. This is why we provide a Simulation Mode.
In Simulation Mode players can test their fighters against different levels of pre-trained AIs. It is a risk-free way of testing a fighter’s readiness before battling in the arena.