1
$\begingroup$
  1   2   3   4   5
|---|---|---|-----|---|---|
|   | > | v |     |   | 5 |
| > | ^ | > | [>] | z | 4 |
| ^ |   |   |  ^  |   | 3 |
| ^ | > | > |  ^  |   | 2 |
| a | ^ |   |     |   | 1 |
|---|---|---|-----|---|---|

Let say we have a GRID like the one above. I want to create score for paths, where every next step has a lower discounted value. F.e. let start with score with range [1,0), step 1:start (1,1) is 0.9, step2:up (1,2) 0.81, step3:up (1,3) 0.73, .... and store the update Score in every cell.

(Later I want to add +/- Rewards along the path)

The first problem is if the path is too many steps the latter steps will have very small values.

Second most important is if paths overlap, the next path will screw the score of the previous path. F.e. cell(4,4)

One solution is for every path to have its own lookup table but this will take too much memory. I want to reuse the grid for all the paths, something like Reinforcement Learning but in the opposite direction, from start to end.

Cant figure the exact formula, has to be something along the line :

V_0 = 1
V_t = decay(V_t-1,%)

sort of !! + reward somewhere

Whats your idea ?

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.