search

zomux / deepy

A highly extensible deep learning framework
MIT License
423 stars 66 forks source link

DQNAgent convergance ? #14

Closed ganarajpr closed 8 years ago

ganarajpr commented 9 years ago

Hi,

I am doing my own experiments with the DQNAgent - the only difference is that I want it to go to the target location and stop.

In contrast, in the puckworld, the agent is continuously looking for new "fruits" to gobble up. In my version, I want it to find one "fruit" in the search space with the most efficient path ( set of actions! ). The agent keeps running without converging on the point that provides the highest reward. Is there anything that I need to change to get this working that way ?

zomux commented 9 years ago

Hi,

Well, There is a hyper-parameter (EPSILON in deepy/experiments/deep_qlearning/agent.py) to control the probability to perform a random action.

You can set EPSILON to zero, and see how it goes.

If it does not work, the network may has some underfitting / overfitting problem, may be some hyper-parameters like "weight_l2" (L2 regularization) should be changed.

Also the object function may also cause the problem, you have to check that.

Please tell me if you find anything about it.