github address

Algorithms

Note: All the base code is from "Hands on Reinfocement Learning".

  • Off-Policy
    • VanillaDQN
    • DoubleDQN
    • DuelingDQN
    • DDPG
    • SACContinuous
  • On-Policy
    • VPG(REINFORCE)
    • ActorCritic
    • TRPO
    • PPO
    • PPOContinuous

Registry

It allows for

  • using decrator to register a new class, and the self.xxx can be overwritten by config.yaml
  • building an intance by Registry().build({"type":"className"}), the dict is the key-value pair for the init() of the given class. "type" is fixed for the className’s key.
  • since we set all the default values in the class’s init(), so Registry().build({"type":"className"}) can write the least number of parameter.
  • if you want to convey param by {"type":"className", "param1":"value1"} this way, remember to comment the coresponding items in the config.yaml, since the value can be overwritten by this file.

Config.yaml

Every class can set param in this file, the file’s priority is the highest.

Wandb

The log system.

Runner

All the supported algorithm has been registried in this file.

Main

The Gallery’s enter point.

logs

打赏作者

发表回复

您的邮箱地址不会被公开。 必填项已用 * 标注

CAPTCHA