Ml-agents: What's the lowest and highest reward?

[lazyvainglory]

I'm trying to fine tune my curriculum, currently I can train my agent to do a task up to 25% of what I want. So I'm trying to find out

1 What is the lowest positive reward and highest negative reward an agent can get by calling AddReward();
2 What is the lowest and highest accumulated reward?
3 Is it possible to know the accumulated reward on a specific agent?

[MinneStephanie2]

Hey,

When you use AddReward than the value has to be between -1 and 1.

[mbaske]

When you use AddReward than the value has to be between -1 and 1.

AFAIK, it is _recommended_ to keep rewards in this range, but training still seems to work fine if I don't. What's the reasoning behind this recommendation?

[mbaske]

Is it possible to know the accumulated reward on a specific agent?

You can call an agent's GetCumulativeReward() method.

[lazyvainglory]

@MinneStephanie2 >

When you use AddReward than the value has to be between -1 and 1.

Let me change the question. What is the highest negative reward and the lowest positive reward. Is it 3 decimal +/- .001 or 6 decimal +/- 000001?

@mbaske Thank you.

[andrewcoh]

There is no bound on what can be passed to AddReward. The range of -1 to 1 is recommended for learning stability purposes.

[MinneStephanie2]

@lazyvainglory

Let me change the question. What is the highest negative reward and the lowest positive reward. Is it 3 decimal +/- .001 or 6 decimal +/- 000001?

Your agent will always try to get a reward as high as posible. You can go as low in that reward as you want.

[lazyvainglory]

@andrewcoh @MinneStephanie2 Thank you