Skip to content

question regarding ADI #1

@Ntiboroo1

Description

@Ntiboroo1

Dear Ktakuya332C,

I'm implementing the DeepCube in Python using TF/Keras.
I would like to ask about the convergence of ADI.

As I understand, the value of a state should reflect the minimum steps needed to solve the state.
My problem with AD is the following:
In the Vx = max(Vxi+A(x,i) ) rule xi can be a state, which is not in the training set. Thus Vxi can be anything, the v = max... rule doesn't apply to it.
Consider a chain with states getting further from the solution:
S0,S1, ..., S10
here: S0 is the solved state, Si needs minimum i rotation to solve it. Let S0,S1,...,S9 being in the training set, and S10 not being in the training set.
Given V(S0) = 0, V(S1) =1, V(S2)=2... V(S10) = 10 is a perfect solution with total loss being 0 for S0...S9. S10 has no neighbour with value 11, but it doesn't matter!

Similarly, V(S0)=0, V(S1) =1, V(S2)=0, V(S3)=-1, V(S4) = -2, V(S5) =-1, V(S6)=0, V(S7)=1 ...V(S10)=4 is a good solution.
In all the cases the value is not useful at all, since it increases instead of decreasing.

How to resolve this problem? My ADI - neural network converges to a solution like above mentioned.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions