Speedup table lookup / interpolation

Question

Jurrien Plijter 2021-8-1

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/890077-speedup-table-lookup-interpolation

编辑： dpb 2021-8-3

Hi, for AI Q-learning I need to interpolate a found state to a table of states to determine a next action.

states_list is a Nx4 matrix (N is in order of 1e6), state is a 1x4 array. I need to find the row (state_index) in the state_list that is the nearest to the state. The function below works, but is far to slow for my goal.

Any suggestions to improve the performance of this function?

function state_index = interpolate_states(states_list, state)
error = abs(states_list(:,1)-state(1))+abs(states_list(:,2)-state(2))+abs(states_list(:,3)-state(3))+abs(states_list(:,4)-state(4));
[~,state_index] = min(error);
end

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

darova 2021-8-1

It can't be faster. You are already using the fastest mthod

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

dpb 2021-8-1

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/890077-speedup-table-lookup-interpolation#answer_758402

编辑：dpb 2021-8-1

在 MATLAB Online 中打开

Slight alteration; didn't try to time it...

[~,ix]=min(sum(abs(states_list-state),2));

is same computationally...

>> x=randi(10,4); y=1:4;                        % some test data
>> e=0;for i=1:4,e=e+abs(x(:,i)-y(i));end       % emulated current solution
>>> all(e==sum(abs(x-y),2))                     % compare to vector solution
ans =
  logical
   1
>>

I wouldn't think it would make a noticeable difference in execution speed and I've never tested it to see if does matter, but you could consider reorienting the data to be by row instead of by column and so are 4 x N and 4x1; this would let sum work in its natural order by column removing the optional second argument.

Another slight overhead reduction might be to write the above inline instead of calling a function for a one-liner...

2 个评论
显示无隐藏无

Jurrien Plijter 2021-8-3

The differences are marginal, but still thank you for your time!

dpb 2021-8-3

编辑：dpb 2021-8-3

Which alternative(s) did you test? The memory orientation one would turn addressing into linear sequence order instead of skipping although since is inside compiled code probably not make a whole lot of difference. Who knows, the JIT may be smart enough to transpose internally, anyway.

请先登录，再进行评论。