This paper proposes a major shift in the decoding of proba- bilistic Tardos traitor tracing code. The goal of the decoder is to accuse colluders but it ignores how they have been mix- ing their copies in order to forge the pirated content. As orig- inally proposed by Tardos, so far proposed decoders are ag- nostic and their performances are stable with respect to this unknown collusion attack. However, this stability automat- ically leads to non-optimality from a detection theory per- spective. This is the reason why this paper proposes to esti- mate the collusion attack in order to approximate the opti- mal matched decoder. This is done iteratively thanks to the application of the well-known Expectation-Maximization al- gorithm. We have dropped the stability: the power of our decoding algorithm deeply depends on the collusion attack. Some attacks are worse than others. However, even for the worst collusion channel, our decoder performs better than the original Tardos decoding.