number in [0, 1], of constraint Ci. A user model provides a partial ordering over all solutions. We will call this the error of the candidate solution which is of the form E(v1,V2→vn)=∑Civi)*w Note that assuming additive independence simplifies the model specification, reducing it to n oft constraints and n weighting coefficients, and simplifies the notion of what a critique is. If the assumption holds, the quality of a candidate solution can be incorrectly computed for only one of two reasons: either the soft constraint for one of the attributes is incorrect, or one of the attributes is weighted improperly. In our implementation in the travel domain, the user is allowed to adjust both of these model parameters directly A user model can either completely or partially describe a user's preferences. In our system the user model initially describes only a few of the user's preferences. As weights are adjusted or constraints are added or updated, the user model becomes a more accurate reflection of the users true preferences 2.4 Candidate/Critique Interaction In this section, we describe the interaction between a candidate/critique agent and a user and specify the input/output behavior of the CCa On each iteration, the CCA uses the current user model to suggest a set of annotated olutions. In our implementation, a solution is annotated if it has the best value in a particular attribute of all the candidate solutions with respect to the current user model. For example, in the travel domain, the cheapest trip of all trips considered by the system would be labeled cheapest". Formally, the CCa is a function from a partial user model and a set of solutions to a small set of suggested solutions. The system calls the CCa with the current user model, and then presents the suggested solutions After a set of solutions has been suggested the user can either choose one and end the interaction, or add a new constraint, modify an existing constraint, or adjust the weighting of a constraint. This can be accomplished, as in our implementation, though the use of a graphical user interface that allows the user to critique the solutions suggested by the CCA. After the user critiques the suggested candidates, the CCa is called with the updated user model, which results n a new set of solutions being suggested 3 CCA Design In this section we first discuss general design principles of a CCA and then describe four general- purpose techniques for building CCA. In the process of describing these techniques, we show how we instantiated them within the travel domain In our implementation, five solutions are suggested in each iteration.number in [0,1], of constraint Ci . A user model provides a partial ordering over all solutions. We will call this the error of the candidate solution, which is of the form E((v1 , v2 , _, vn )) = i = 1 n ∑ Ci (vi )*wi Note that assuming additive independence simplifies the model specification, reducing it to n soft constraints and n weighting coefficients, and simplifies the notion of what a critique is. If the assumption holds, the quality of a candidate solution can be incorrectly computed for only one of two reasons: either the soft constraint for one of the attributes is incorrect, or one of the attributes is weighted improperly. In our implementation in the travel domain, the user is allowed to adjust both of these model parameters directly. A user model can either completely or partially describe a user’s preferences. In our system, the user model initially describes only a few of the user’s preferences. As weights are adjusted or constraints are added or updated, the user model becomes a more accurate reflection of the user’s true preferences. 2.4 Candidate/Critique Interaction In this section, we describe the interaction between a candidate/critique agent and a user and specify the input/output behavior of the CCA. On each iteration, the CCA uses the current user model to suggest a set of annotated solutions. In our implementation, a solution is annotated if it has the best value in a particular attribute of all the candidate solutions with respect to the current user model. For example, in the travel domain, the cheapest trip of all trips considered by the system would be labeled as “cheapest”. Formally, the CCA is a function from a partial user model and a set of solutions to a small set 3 of suggested solutions. The system calls the CCA with the current user model, and then presents the suggested solutions. After a set of solutions has been suggested the user can either choose one and end the interaction, or add a new constraint, modify an existing constraint, or adjust the weighting of a constraint. This can be accomplished, as in our implementation, though the use of a graphical user interface that allows the user to critique the solutions suggested by the CCA. After the user critiques the suggested candidates, the CCA is called with the updated user model, which results in a new set of solutions being suggested. 3 CCA Design In this section we first discuss general design principles of a CCA and then describe four general￾purpose techniques for building CCA. In the process of describing these techniques, we show how we instantiated them within the travel domain. ——— 3 ———— In our implementation, five solutions are suggested in each iteration