I started to learn reinforcement learning a few days ago. And I want to use that to solve resource allocation problem something like given a constant number, find the best way to divide it into several real numbers each is non-negative.

For example, to divide the number 1 into 3 real numbers, the allocation can be:

[0.2, 0.7, 0.1]

[0.95, 0.05, 0] ...

I do not know how to represent the action space because each allocation is 3-dimensional and each dimension is real-valued and each other correlated.

In actor-critic architecture, is it possible to have 3 outputs activated by softmax in the actor's network each represents one dimension in the allocation?


There is a playlist of videos. A user can switch to the next video at any time. More buffer leads to better viewing experience but more bandwidth loss if user switches to the next video. I want to optimize the smoothness of playback with minimal bandwidth loss. At each time step, the agent decides the bandwidth allocation to download current video and the next 2 videos. So I guess the state will be the bandwidth, user's behavior and the player situation.

  • 1
    $\begingroup$ Does your problem naturally have states, time steps and sequences of actions? Without these things, framing it as a reinforcement learning problem can be counter-productive. So it might be worth adding if your primary goal is to solve your original problem (and maybe explain it some more, as you are stuck on representation), or to study RL? $\endgroup$ – Neil Slater Jun 2 '19 at 15:06
  • $\begingroup$ What is the goal? If you just want to divide into 3 random parts it is an easy problem. What is a good result? $\endgroup$ – Miguel Saraiva Jun 2 '19 at 15:24
  • $\begingroup$ I have edited the problem to make it clearer. $\endgroup$ – AlanRivers Jun 2 '19 at 15:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.