Noise parameters in Reinforcement learning DDPG

What should be the values of Noise parameters (for agent) if my action range is between -0.5 to -5 in DDPG reinforcement learning I want to explore whole action range for each sample time? Also is there anyway to make the noise options (for agent) independent of sample time?

Answer 1

[Drew Davis]

Hi Surya

It is fairly common to have Variance*sqrt(SampleTime) somewhere between 1 and 10% of your action range for Ornstein Uhlenbeck (OU) action noise. So in your case, the variance can be set between 4.5*0.01/sqrt(SampleTime) and 4.5*0.10/sqrt(SampleTime). The other important factor is the VarianceDecayRate, which will dictate how fast the variance will decay. You can calculate how many samples it will take for your variance to be halved by this simple formula:

halflife = log(0.5)/log(1-VarianceDecayRate)

It is critically important for your agent to explore while learning so keeping the VarianceDecayRate small (or even zero) is a good idea. The other noise parameters can usually be left as default.

You can check out this pendulum example which does a pretty good job of exploring during training.

The sample time of the noise options will be inherited by the agent, so it is not necessary to configure. By default, the noise model will be queried at the same rate as the agent.

Hope this helps

Drew

[Drew Davis]

I realized I read your range wrong, I initially thought it was -0.5 to 0.5. Edited the above answer

[Surya teja Tunuguntla]

Thank you Drew for your suggestions, but I used tanh layer at the end of actor network and mapped the values from range [-1,1] to [-5,-0.5] using linear mapping and it worked fine. I used variance of 0.15/sqrt(sampletime) and variance decay rate 1e-6 for the above mentioned model. Nevertheless I will try your suggested method as well.

And to avoid algebraic loops, I used 1 timestep lag block, hope it doesn't affect the model.

[Rajesh Siraskar]

Dear Drew, Many thanks for the half-life formula.

Could you please point me to a source for this formula, as I need to reference it.

Although my DDPG training is still under tuning, I think it may have helped me tune it.

[Drew Davis]

You can derive this formula pretty easily:

decayfactor = 0.5 = (1 - decayrate)^(#steps)

[Rajesh Siraskar]

Thank you Drew