ML-agent supports two types of learning:
1.RL:Learns by getting rewards.
2.Imitation:Learns by imitatating what the player does.
Imitation learning is how you teach your ai directly how to behave in order to achieve a certain goal.
Firstly,set up the scene with some randomness,so that the ai doesn't know how to solve just one specific set of positions,so we randomize
Two types of imitation learning that you can use [GAIL,BC]
BC - Behavior Cloning//行为克隆是最简单的模仿学习。
GAIL-Generative Adversarial Imitation Learning,the goal of the discriminator is to figure out if a certain action came from the agent or from the demo.so essentially over time our agent will learn how to behave like the demo in order to trick the discriminator.
GAIL works by trying to trick a discriminator into pretending that the actions came from the demo,whereas BC simply tries to copy exactly what you did ,the limitation of the BC is that it can never get better than the demos.So in order to get the best results,we need to combine all three.
First use BC ,it learns to act exactly like you,then when combined with GAIL,it learns to act similarily to you while achieving the same goal,and when combined with extrinsic rewards,it continues improving upon those two,that's how we get superhuman learning .
Visualization in tensorboard
Personally Speaking:
I have watched some basic project completed by ml-agents.Personally speaking,it feels like
Reference
【Unity 教学】【中字】教你的人工智能！使用 Unity ML-Agents进行模仿学习！_哔哩哔哩_bilibili