There are no semi-weak keys either. Semi-weak keys come in pairs, where the second key decrypts the first. DES has 16 of them, including the four weak keys.
A weak or semi-weak key occurs if there is another key that generates an identical key schedule, but in the reverse order. These keys can be found by setting up a series of linear (under XOR) equations expressing the fact that the schedule of key 1 is the reverse of the schedule of key 2, then solving the equations. The number of independent variables in the solution gives the log base 2 of the number of weak keys.
For ICE, there were 960 equations (16 rounds, 60 bits per round) and 129 variables (2 x 64-bit keys, plus an inversion bit). The solution was "1=0", which means that there are no keys that satisfy the equations.
They are caused in DES by the fact that key bits are only used with the XOR function. If both key and plaintext bits are inverted, the inversions are cancelled out by the XOR function, and DES behaves linearly. However, ICE also uses key bits to permute the outputs of the E boxes, so if the key is inverted, the S-boxes will receive totally different inputs.
The use of keyed permutation after the E boxes means that an attacker cannot know which S-box will be affected by a particular input bit. However, since the keyed permutation acts to simply swap bits between the left and right halves of a 32-bit value, the attacker can use inputs whose leftmost 16 bits are the same as the rightmost 16 bits. This enables the attacker to send known differences to the S-boxes, but it usually also means that twice as many S-boxes have to be attacked simultaneously, often with low-probability differences. This typically at least squares the number of pairs required to achieve a result.
The best input differences for attacking the ICE F-function are b2d6b2d6 and cad6cad6, both of which produce a zero output difference with probability 4320/2^40. They can be combined into 5-round characteristics which have a probability of 2^-55.85, and it is these characteristics that may be able to break Thin-ICE, which is an 8-round cipher.
The resistance of ICE to linear cryptanalysis is due to the larger S-boxes, and to the keyed permutation, which roughly squares the effort otherwise required.
ICE avoids this weakness in its extended variants by extending the key schedule with insertions in the middle of the schedule. Although ICE-n effectively encrypts the data n times with n different 64-bit keys, it does this not by encrypting with one key after another, but by doing half encryptions (i.e. the first 8 rounds) n times, then doing the second halves n times.
So it must be remembered that although the strength of ICE-n under ciphertext-only attacks is probably 2^64n, the strength of all ICE variants under chosen-plaintext is, at best, 2^64.