This paper proposes a model for color perception which accounts for variations in the dimension of the space of perceived colors. The model assumes that there is only one type of cone with only one shape of response curve, but that different cone's response curves differ by translation. It also assumes that the final discrimination system, learned from originally random connections, maximizes discrimination in the normal visual environment. Learning mechanisms are discussed, and the form which the final discrimination system ought to take is plausibly derived. An algorithm for the tristimulus curves is obtained from this model, and it is shown that a good fit of the empirical data can be obtained.