Reinforcement Learning

An paradigm of ML: How intelligent agent should take actions to accumulate maximum reward

Given an action by the agent,

• Environment changes state
• Environment gives reward
• Agent can take next action based on knowledge
• Physically based

UnityML

Agent:

• Collection of observations from environment
• Action execution

Brain:

• Decision making for linked agents
• Academy to track iterations, set simulation speed and reset environment.

Academy:

• Tracks iterations
• Global environment variables

Proximal Policy Optimization:

• Used to train advanced models

Model Locomotion

• Force applied is based on acceleration $F=ma$
• each limb being subjected to a force with the limb’s mass
• $F = (v_1 - v_0)m/t = ma$