I assume here that your plaintexts and ciphertexts P, P* and C are all 128-bit blocks.
If your keys k and k* have 128-bit length (i.e. you are using AES-128), then, with probability about 36.8%, there is no solution: more formally, if you consider the set of all possible values for C and P* (2256 combinations), then for about e-1 of them there is no k* such that AES_k*(P*) = C.
This follows from the idea that, for a given value of P*, the function which transforms k* into AES_k*(P*) should behave as a random function, and a random function from a set of size N to a set of identical size N has average coverage of 1-e-1 of the destination set. Here, for a given P*, there are about 63.2% of 128-bit words which are possible outputs of AES encryption of P* with a 128-bit key.
On the other hand, if you allow k* to be wider (AES also accepts 192-bit and 256-bit keys) then there should be a great many solutions k* to your equation.
Anyway, actually finding k* (even a 192-bit or 256-bit k*) should be infeasible, with work factor close to 2128 operations. Being able to find k* with less work than that could be viewed as a structural flaw in AES. Knowledge of P and k helps in no way: for the given 128-bit ciphertext C, it is easy to find matching pairs (P, k).
Note: if you take AES and swap the roles of the plaintext and the key, then you get a crude emulation of a hash function with limited input, and 128-bit output. What you are asking for is the feasibility of a preimage attack on that hash function.