The aim of good enough answers in the OpenKnoweldge project is to facilitate the setting up of interactions such that the participants in it are likely to be satisfied with the outcome. In practice this means that the peers assigned to each role in the interaction will perform their role with a reasonably high degree of competency.
In order to perform a role competently, a peer needs to be both willing and able to do so.
1) Being willing to means that the peer will do its best to satisfy its constraints correctly and to be online throughout the interaction: it will not malicously fail to perform its role.
2) Ability is a two-fold issue.
- Firstly, the peer must be able to adequately satisfy the constraints on its role, which means that it must be able to satisfactorily map them into its own methods. As explained in the section on matching, this may not be a perfect match: sometimes these matches are approximate. The ability of the peer to satisfy a constraint is proportional to the matching score it obtained for matching that constraint to one of its methods.
- Secondly, the services or information provides must be of a reasonable quality. For example, if the peer is acting as a car seller then it must not only be able to match the correct constraints on the role but the car, when it is deliverered, must also be of a high quality.
A good enough score is there a value in [0 1] that estimates, based on these three issues, how well we believe a peer assigned to a role will perform that role.
This score is difficult to estimate, since it contains a lot of unknowns. The only aspect that can be directly calculated is the SPSM - though even this is not certain, as this is calculated by the peer wishing to play a role and cannot be replicated by any other peer as that peer's methods are private; therefore there is no way of verifying that the reported score is accurate.
For the other two aspects, we must depend on past experience to form some estimate of behaviour: this is what is done by the trust module. Trust assigns a score for a peer playing a particular role in a particular interaction, after that interaction is completed. If a peer persistently scores highly, we can guess that it is generally willing to perform its role well. If a peer scores highly on roles that are the same or very similar to the role it is currently playing, we can guess that is able to perform that role with high quality - e.g., a car seller will provide good quality cars. An important part of the of the trust model is therefore using the SPSM algorithm to assess whether roles are similar by calculating how similar their constraints are. If there is not enough direct experience available - is will often be the case - this information can be shared through gossiping with other peers. Further details of the trust model - what its aims are and how it works - can be found on the trust research page.
The GEA score for a peer playing a role is therefore calculated as:
for a role with i constraints, where trust_score is the trust score for that peer playing that role in that IM.
Although the matching score must be taken on trust, a peer that repeatedly reports a higher than accurate matching score will usually fail to perform to the expected level of satisfaction and its trust score will therefore go down. Overall its GEA score will thus be reduced, and so it is not in a peer's interests to frequently lie about their scores.
Further information and algorithms for GEA can be found in Deliverable 4.5.
GEAs have been thoroughly evaluated within the e-response test bed, with promising results. These are reported in Deliverable 4.9.