-
Notifications
You must be signed in to change notification settings - Fork 0
/
note.txt
19 lines (17 loc) · 1005 Bytes
/
note.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Optimal Action-value and State-value functions
If the entire environment is known, such that we know our reward function and transition probability function, then we can solve for the optimal action-value and state-value functions via Dynamic Programming like
Policy evaluation, policy improvement, and policy iteration
However, typically we don't know the environment entirely then there is not closed form solution in getting optimal action-value and state-value functions. Hence, we need other iterative approaches like
Monte-Carlo methods
Temporal difference learning (model-free and learns with episodes)
On-policy TD: SARSA
Off-policy TD: Q-Learning and Deep Q-Learning (DQN)
Policy gradient
REINFORCE
Actor-Critic
A2C/A3C
ACKTR
PPO
DPG
DDPG (DQN + DPG)
https://towardsdatascience.com/how-to-install-openai-gym-in-a-windows-environment-338969e24d30