[MLS-C01] [Algorithms] Reinforcement Learning Algorithms

Defined

Machine learning technique which learns a strategy, called a policy, that optimizes an objective for an agent acting in an environment
Agent performs an action, checks the state of the environment, and gets rewarded based on the resulting state of the environment
Attempts to achieve the goal of maximizing the long-term reward that the agent receives resulting from its actions
Good for solving problems where you want your model to make independent decisions.
Markov Decision Process (MDP)
- Based on Markov Decision Processes (MDPs) models
- Works through a sequence of time steps with each step containing the following
  - Environment: operating space of the RL model
  - State: complete information describing the environment and relevant past steps
  - Action: agent’s activity
  - Reward: number that reflects the state resulting from the last action
  - Observation: data about the state of the environment available at each step

Reinforcement Learning in TensorFlow and Apache MXNet
- Uses a Reinforcement Learning toolkit
  - Manage the agents interaction with the environment
  - SageMaker supports Intel Coach and Ray RLlib toolkits
- Uses an environment
  - Custom environment
  - Open source environments: EnergyPlus, RoboSchool
  - Commercial environments: MATLAB, Simulink
Important Hyperparameters
1. learning_rate: how fast the model learns
2. discount_factor: take action with short-term or long-term rewards
3. entropy: degree of uncertainty; exploit what’s already known versus thorough exploration of the environment

本文由 Oscaner 创作, 采用知识共享署名4.0 国际许可协议进行许可
本站文章除注明转载/出处外, 均为本站原创或翻译, 转载前请务必署名