There are drawbacks, like requirement of knowing the dynamics of the environment, which we face when using techniques like Value Iteration and Policy iteration. To overcome this drawback we can use sampling-based approximation techniques like Q-Learning.
In Q-Learning we:
- Replace Bellman equation (written in iteration format):
- And write it in the form of Expectation:
- Now for a single state action (s, a) pair we can replace expectation by samples:
- For a single state-action pair (s, a), receive s' ~ P(s'|s,a)
- Consider our old estimate:
- Consider our new estimate:
- Incorporate the new estimate into a running average:
The task is to maximize a reward in a world that consists of an agent that can navigate in 4 directions - North, South, East and West. With a 20% of equally likely chance of deviating to left or right from the action asked to perform.
 |
|---|
| Src: UC Berkley 2017 Deep RL bootcamp Lecture 1 slides |
Modify main.json to suit your needs. The key names are self explanatory. Then run python main.py.
You can also create your own <user-defined>.json file with every paramter defined and then run python main.py --json_path <user-defined>.json