Problem Definition

	The basic idea behind P2P systems is that peers interact with other peers in order to fulfil their goals.
	Translated to the OK system, this implies that a peer will have to choose an appropriate interaction model for fulfilling its goal, along with the appropriate peer for playing each of the chosen interaction model's roles.
	Trust, in the OK system, is defined as the problem of choosing appropriate peers for playing a specific role in a given interaction model.
	In summary, after choosing a given interaction model, the peer will have to select the appropriate peers for filling in each of the roles of this interaction. This is where trust comes in.

Methodology

The methodology followed in addressing the issue of trust is summarised by the algorithm presented below. For additional explanation, click on a step of the algorithm for its corresponding description.


Description

1. The goal is to decide whether P is appropriate for playing a given role. Φ can be as general as playing the role or as specific as performing a specific and crucial action required by that role, such as paying on time.
2. Peers announce their capabilities of playing a given role by posting a matching score MS. This score describes whether the peer can fulfil the role's constraints by matching the peer's OKCs to the constraints of the given role. Deliverable 4.4 describes the matching algorithm in detail. However, peers may lie about their capabilities. Hence, in addition to the matching score, a trust score needs to be computed, describing how much is the peer trusted in performing Φ. This is what the remainder of this algorithm tries to achieve.
3. Past experiences are used to construct a probability distribution describing the possible outcomes of future commitments. The final trust measure is based on this distribution. Hence, the state space of this probability distribution needs to be defined by computing the terms similar to Φ. For example, if the buyer orders wine from a wine store, it might be possible to receive (either by mistake or on purpose) wine, beer, whiskey, etc. But it will be highly unlikely that he will receive the latest Harry Potter novel for instance.
4. Given the current commitment to performing Φ, only past experiences whose commitments were similar to Φ can provide an insight to what the possible outcome of committing to Φ will be. Hence, only similar past experiences are selected for the computation of trust.
5. The probability distribution describing the possible outcomes of committing to Φ is updated by considering each past experience ψ, one at a time.
6. Given a single past experience, the probability distribution describing the possible outcomes of committing to Φ is updated by updating the probability of each of the elements of the distribution's state space SP. Hence, the elements of SP are studied one at a time, in order to update the probability of each with respect to the chosen past experience.
7. This is an AND operator. To compute the probability of peer P performing Φ', given its commitment to Φ, both the capability and willingness of the peer should be evaluated by consulting past experiences. This is because the matching score MS made public by peer P, and describing the peer's capability, cannot be fully trusted.
8. If the peer has succeded in performing Φ' at least once in the past, then the peer is said to be capable of performing this action. Hence, to compute the probability of the peer being capable of performing Φ', the entire data base of past experiences is searched for all previously successful actions. The successful action that is the most similar to Φ' is obtained. The probability of being capable of performing Φ' is then defined as this maximum similarity measure.
9. To compute the probability of peer P performing Φ', given its commitment to Φ, one should investigate the outcome of all past experiences whose commitments were similar to Φ. The higher the similarity measure between Φ' and the outcomes of those selected past experiences, then the higher the probability of having Φ' performed.
10. After computing the probability of being capable of performing Φ' (P_can(Φ')) and the probability of actually performing Φ' given that the peer has committed to Φ (P_does(Φ'|Φ)), the final probability measure is defined as the product of the two.
11. As previously illustrated, the probability distribution is obtained by computing the probability of each element of the state space SP; hence, an iteration over these elements is needed.
12. As previously illustrated, the probability distribution is updated only after considering all similar past experiences; hence, an iteration over these experiences is needed.
13. As previously illustrated, the probability distribution is in fact described by the probabilities of each element of the distribution's state space.
14. With time, old experiences lose their effect, as they become less and less relevant. Hence, a time decay factor needs to be applied to the probability distribution. Of course, the measure of old is relative and context dependent. Hence, this measure may only be specified by the peer performing the trust computation.
15. Given the probability distribution describing the possible outcome of peer P committing to Φ, the trust measure describing how much is peer P trusted in performing Φ depends on the peer and the context. For example, if a peer orders a bottle of wine, the peer will only be satisfied if it receives that exact bottle of wine. However, another peer might also be satisfied if it receives whiskey instead. Hence, the computation of trust is a subjective matter. Deliverable 4.5 provides several methods for computing trust by allowing the peer to specify its ideal enactments, its preferred enactments, or its satisfactory ones.
16. After the peer obtains both the matching score MS and the trust score TS, both scores are aggregated following one of the algorithms described by Deliverable 4.5 in order to achieve one final score. This final score describes the appropriateness of peer P in performing action Φ.

Further Information

• Deliverable 4.7 provides a general overview of available trust models for multigent systems.
• Deliverable 4.5 presents our proposed methodology for dealing with trust in the OpenKnowledge system.
  
• Deliverable 4.8 illustrates the implementation of the proposed trust module as an OK plug-in component.
• Deliverable 4.9
• This document provides additional information on how the issue of trust may be dealt with in an automated manner by addressing it from within a running LCC interaction.