5.1 A NÁLISIS DE VULNERABILIDAD SOCIO AMBIENTAL
5.1.1 Cuenca del Río Gacheta
5.1.1.1 Subcuenca Río Amoladero
This section provides an agent-based VO scenario in which we demonstrate the use of TRAVOS. We begin by stating that there is a need to create a VO to meet a specific requirement to provide a composite multimedia communication service to an end user. This consists of the following basic services: text messaging, multimedia streaming, HTML content provision, and phone calls (this example is adapted from one given by Norman et al. (2003)). Now, assume agent a1 has identified this need and wishes to
capitalise on the market niche. However, a1 only has the capability to provide a text
messaging service, and can only achieve its goal by forming a VO with an agent that can supply a service for phone calls and one for HTML content. For simplicity, we assume that each agent in the system has the ability to provide only one service. Agent a1 is
aware of three agents that can provide a phone call service, and its interaction history with these is shown in Table 4.2. Similarly, it is aware of three agents that are capable of providing HTML content, and its past interactions with these entities are given in Table 4.3. We also assume that a truster’s prior parameter distribution for all agents is
Agent Past interactions Successful Unsuccessful
a2 17 5
a3 2 15
a4 18 5
Table 4.2: Agenta1’s interaction history with phone call service provider agents.
Agent Past interactions
Successful Unsuccessful
a5 9 14
a6 3 0
a7 18 11
Table 4.3: Agenta1’s interaction history with HTML content service provider agents.
uniform:
αprior= 1, βprior= 1
Agent a1 would like to choose the most trustworthy phone call and HTML content
service provider from the selection. The following describes how this is achieved using TRAVOS.
4.5.2.1 Calculating Trust
Using the information from Tables 4.2 and 4.3,a1can determine the number of successful
interactions, m, and the number of unsuccessful interactions, n, for each agent it has interacted with. Feeding these into Equations 4.9 and 4.10, a1 can obtain a parameter
distribution which summarises each agent’s likely behaviour in future interactions; for example, the shape parameters α andβ, for a2, are calculated as follows:
Using Table 4.2: ma1,a2 = 17, na1,a2 = 5.
Using Equations 4.9 & 4.10: α= 17 + 1 = 18 and β = 5 + 1 = 6.
The hyperparameter for each agent is then used to estimate the probability that each agent will cooperate in any future interaction. In line with Section 4.1, we calculate this estimate as the expected value of the parameter distribution (Equation 4.5); for example, the estimate, ϑa1,a2, fora2 is calculated as follows:
Using Equation 4.5: ϑa1,a2 =
α α+β =
18
18+6 = 0.75.
The above estimate givesa1 an assessment ofa2’s likely behaviour based on direct inter-
actions. However, as discussed in Section 4.4, a1 may wish to determine if the accuracy
of this estimate is sufficient to avoid the need to gather reputation. To do this, we calculate the posterior probability that the true value forθa1,a2 lies within an acceptable
Agent α β ϑa1,ax γa1,ax a2 18 6 0.75 0.9806 a3 3 16 0.1579 0.9798 a4 19 6 0.76 0.9835 a5 10 15 0.4 0.9657 a6 4 1 0.8 0.8704 a7 19 12 0.6129 0.9822
Table 4.4: Agent a1’s calculated trust and associated confidence level for HTML
content and phone call service provider agents.
margin of error around the estimate. We can calculate this using the parameter distri- bution as follows. First, we decide on an acceptable error margin,ϑa1a2±, whereis a
suitable value, such as 0.2. Second, we integrate the parameter distribution over the area defined by the error margin. Finally, we determine some threshold for this probability, above which the estimate gives an acceptable level of accuracy; for example, we could define a thresholdτ as 0.95. The proceeding example illustrates this calculation fora1’s
estimate for a2, using = 0.2; we denote the resulting confidence value asγa1,a2:
γa1,a2 = Rϑa1,a2− ϑa1,a2+ B α−1(1−B)β−1dB R1 0 Uα−1(1−U)β−1dU = R0.55 0.95 Bα −1(1−B)β−1dB R1 0 Uα−1(1−U)β−1dU = 0.98
The hyperparameters, estimate and associated confidence for each agent,a2 toa7, which
a1computes using TRAVOS, are shown in Table 4.4. From this, it is clear that the trust
values for agents a2,a3 and a4, all have a confidence aboveτ (=0.95). This means that
a1 does not need to consider the opinions of others for these three agents. Agenta1 is
able to decide thata4is the most trustworthy out of the three phone call service provider
agents and chooses it to provide the phone call service for the VO.
4.5.2.2 Calculating Reputation
The process of selecting the most trustworthy HTML content service provider is not as straightforward. Agent a1 has calculated that out of the possible HTML service
providers,a6 has the highest trust value. However, it has determined that the confidence
it is willing to place in this value is 0.8704, which is below that ofτ and means thata1
has not yet interacted with a6 enough times to calculate a sufficiently confident trust
value. In this case, a1 has to use the opinions from other agents that have interacted
with a6, and form a reputation value for a6 that it can compare to the trust values it
has calculated for other HTML providers (a5 and a7).
Suppose that a1 is aware of three agents that have interacted with a6, denoted by a8,
a9 anda10, whose opinions abouta6 are (15,46), (4,1) and (3,0) respectively. Agenta1
Opinions from providers: a8 = (15,46), a9= (4,1) and a10= (3,0)
Using Equations 4.19 & 4.20: M = 15 + 4 + 3 = 22, N = 46 + 1 + 0 = 47
Using Equations 4.17 & 4.18: α= 22 + 1 = 23, β= 47 + 1 = 48
Having obtained the shape parameters,a1 can obtain an estimate for a6 using Equation
4.5, as follows: Using Equation 4.5: ϑa1,a6 = α α+β = 23 23+48 = 0.3239
Now, a1 is able to compare the trust in agents a5, a6 and a7. Before calculating the
trustworthiness of a6, agent a1 considered a6 to be the most trustworthy (see Table
4.4). Having calculated a new trust value for agent a6 (which is lower than the first
assessment), agent a1 now regardsa7 as the most trustworthy. Therefore a1 choosesa7
as the service provider for the HTML content service.
4.5.2.3 Handling Inaccurate Opinions
The methoda1 uses to assess the trustworthiness ofa6, as described in Section 4.5.2.2,
is susceptible to errors caused by reputation providers giving inaccurate information. In our scenario, suppose a8 provides the HTML content service too, and is in direct
competition witha6. Agenta1 is not aware of this fact, which makesa1 unaware thata8
may provide inaccurate information abouta6 to influence its decision on which HTML
content provider agent to incorporate into the VO. If we examine the opinions provided by agents a8, a9 and a10, which are (15,46), (4,1) and (3,0) respectively, we can see
that the opinion provided by a8 does not correlate with the other two. Agents a9 and
a10provide a positive opinion of a6, whereas agenta8 provides a very negative opinion.
Suppose thata8 is providing an inaccurate account of its experiences with a6. We can
use the mechanism discussed in Section 4.3 to allow a1 to cope with this inaccurate
information, and arrive at a better decision that is not influenced by self-interested reputation providing agents (such as a8).
Before we show how TRAVOS can be used to handle such inaccurate information, we must assume the following. Agent a1 obtained reputation information from a8, a9 and
a10on several occasions, and each timea1recorded the opinion provided by a reputation
provider and the actual observed outcome (from the interaction with an agent to which the opinion is applied). Each time an opinion is provided, the outcome observed is recorded by updating a frequency bin corresponding to the interval, ΘCr, which the received opinion belongs to. Agent a1 maintains information of like opinions in bins as
shown in Table 4.6. For example, ifa8 provides an opinion that is used to obtain a trust
value of 0.254, then the actual observed outcome (successful or unsuccessful) is stored in the 0.2< E[θatr,ate|φr]≤0.4 bin.
Agent Weighting Adjusted Values
µ σ α β
a8 0.0049 0.4988 0.2875 1.0095 1.0144
a9 0.7802 0.6672 0.1881 3.5215 1.7567
a10 0.7424 0.7227 0.1956 3.0629 1.1751
Table 4.5: Agenta1’s adjusted values for opinions provided bya8,a9anda10.
[0,0.2] [0.2,0.4] [0.4,0.6] [0.6,0.8] [0.8,1] Total
m n m n m n m n m n
a8 2 0 11 4 0 0 0 0 2 3 22
a9 0 2 1 3 0 0 22 10 6 4 48
a10 1 3 0 2 0 0 18 8 5 3 40
Table 4.6: Observations made bya1 given opinions from reputation sources. m rep-
resents that the interaction (to which the opinion applied) was successful, and likewise n means unsuccessful.
Using the information shown in Table 4.6, agent a1 can calculate the weighting to
be applied to the opinions from the three reputation sources by applying the tech- nique described in Section 4.3. In so doing, agent a1 uses the information from the
bin that contains the opinion provided, and integrates the beta distribution between the limits defined by the bin’s boundary. For example, a8’s opinion falls under the
0.2< E[θatr,ate|φr]≤0.4 bin. In this bin, agenta1 has recorded thatm= 11 andn= 4. Thesemandnvalues are used to obtain a beta distribution,d(θatr,ate|φo), which is then integrated between 0.2 and 0.4 to give a probability of accuracyρa1,a8 = 0.0049 fora6’s
opinion. Then, by using Equations 4.22 and 4.23, agent a1 can calculate the adjusted
mean and standard deviation of the opinion, which in turn gives the adjustedα and β
parameters for that opinion. The results from these calculations are shown in Table 4.5. Summing the adjusted values for αand β from Table 4.5,a1 can obtain a more reliable
value for the trustworthiness ofa6. Using Equation 4.5,a1calculates an estimateϑa1,a6 =
0.7419 fora6. This means that from the possible HTML content providers,a1 now sees
a6 as the most trustworthy and selects it to be a partner in the VO. Unlikea1’s decision
in Section 4.5.2.2 (whena7 was chosen as the VO partner), here we have shown how a
reputation provider cannot influence the decision made by a1 by providing inaccurate
information.