The implementation is based scenario where agents need to agree on a plan to attend a conference in Paris travelling from Inverness. This example treats multi-agent plans in a different way as the way they were printed in Chapter 3. Agents execute the same plan but still they may have different beliefs about its specification. The situation is the following: two agents (named John and Paul) both have different plans to reach Paris and different beliefs about the current circumstances.
The actions to travel are: travel by train takeT rain(), take a flight takeF light() and take a coach takeCoach() through four cities Inverness, Manchester, London and Paris. I assume that both agents execute the same action i.e. John takes a train and Paul takes a train. The actions are the following:
• takeCoach(Inverness, M anchester)
• takeCoach(M anchester, London)
• takeCoach(London, P aris)
• takeT rain(Inverness, M anchester)
• takeT rain(M anchester, P aris),
• takeF light(Inverness, P aris)
• takeF light(London, P aris)
Each city has restrictions on the availability of the train station, airport and coach station.The values that can be associated with the plans are:
• Valuev1 =money, the cheapest option.
• Valuev2 =time, the fastest option.
• Valuev3 =f riendship, agents travelling together.
• Valuev4 =comf ort, the most comfortable way to travel.
Figure 6.6 presents the available actions that agents can choose with the restrictions on each city. Table 6.6 presents some the possible plans agents can propose. The following sections focus on the problem of creating a proposal taking into account all the details that the problem poses and how these proposals could be challenged.
Inverness
Paris London Manchester
Coach Train Flight
Table 6.6: Agents’ plans for the Journey Example
Plan Actions Values
P L1 - Coach takeCoach(Inverness, M anchester) v1 =money + takeCoach(M anchester, London) v2 =duration − takeCoach(London, P aris) v3 =f riendship = v4 =comf ort −
P L2- Trains takeT rain(Inverness, M anchester) v1 =money = takeT rain(M anchester, P aris) v2 =duration = v3 =f riendship + v4 =comf ort +
P L3 - Flight takeF light(Inverness, P aris) v1 =money − v2 =duration + v3 =f riendship − v4 =comf ort =
P L4 - Coach-Train takeCoach(Inverness, M anchester) v1 =money = takeT rain(M anchester, P aris) v2 =duration − v3 =f riendship + v4 =comf ort =
P L5 - Train-Flight takeT rain(Inverness, London) v1 =money − takeF light(London, P aris) v2 =duration = v3 =f riendship = v4 =comf ort =
P L6 - Coach-Train- takeCoach(Inverness, M anchester) v1 =money − Flight takeT rain(M anchester, London) v2 =duration = takeT rain(London, P aris) v3 =f riendship = v4 =comf ort −
The test cases are formed by providing the agents with:
1. A set of propositions that represent constraints of the “social context”. 2. An initial view of the world.
3. A set of action specifications. 4. A set of plans.
5. A set of values.
6. A preference order over values.
Agents’ plans are presented in Table 6.6 together with the status of the values (pro- moted (+), demoted (−) or neutral (=)). Actions are meant to be executed simultane- ously by both agents. Although each individual action could be associated with a value, for the sake of simplicity here I will only consider values related to the plan as a whole. The validity of some elements in the plans is changed for the different test cases and/or world representation for each agent to create different scenarios. The validity of elements (actions conditions, action effects, norms) is represented using a “token attribute” associated with each element. That the “token” is f alserepresents that the element validity against the context has expired. The evaluation of arguments that attack the proposal is done automatically through the token functionality.
Agents are pre-loaded four different sets of information about the world and plans (presented in Table 6.7). These sets are then combine them with five different preference orders for the agents to generate twenty test cases. In Table 6.7, a check mark (X) indicates the validity of the element and a cross (×) indicates that an invalid element exists in the specification. The twenty test cases constructed with information in Tables 6.7 and 6.8 to generate dialogue runs. The test cases are described below:
• Test case A: All the plans and information about the world is valid for both agents. Only questions suggesting alternative plans are generated for this test case. Agents focus on selecting the best option on each run.
• Test case B: John’s plansP L1, P L2, P L3 are invalid inducing questions about the validity and possibility of the action elements. Any preference over these plans has to be re-evaluated after the dialogue. The social constraints and initial state of Paul are not valid and generate possibility questions for norms and suitability questions regarding the initial state.
• Test case C: John’s plans P L2,P L3 are invalid. Again any preference over these plans has to be re-evaluated after the dialogue. Paul’s plan P L4 is also invalid.
Table 6.7: Test case specifications
Test John’s John’s Paul’s Paul’s
case Plans Beliefs Plans Beliefs
P L1 X Social constraints X P L4 X Social constraints X A P L2 X Initial state X P L5 X Initial state X P L3 X Action specification X Action specification X P L6 X
P L1 × Social constraints X P L4 X Social constraints × B P L2 × Initial state X P L5 X Initial state × P L3 × Action specification X Action specification X P L6 X
P L1 X Social constraints X P L5 X Social constraints X C P L2 × Initial state X P L4 × Initial state X P L3 × Action specification X Action specification X P L6 X
P L1 X Social constraints X P L4 X Social constraints X D P L2 × Initial state × P L5 X Initial state X P L3 X Action specification X Action specification X P L6 X
• Test case D: John’s initial state is not valid nor is his plan P L2. This case aims to question the proposals mainly at the plan suitability level.
Table 6.8 presents the five sets of preference orders used for both agents. Agents put forward first their preferred plans and proceed in descending order of priority. Agents may change their preferred plan during the dialogue depending on the outcome of the questioning process. More details on how the agents’ preferred plan changes after the dialogue are given in the results presented in the next chapter. Each test case is executed through a dialogue simulation six times to test the different strategies used (the two strategies from Chapter 5 and a random approach) and the agent that starts the dialogue (two agents), since which agent goes first might have a significant influence on the dialogue. Table 5.2 in the previous chapter presents the order in which questions are considered for both strategies.
Table 6.8: Test cases preferences and the preferred plan for both agents.
John preference Paul preference
Test Value Preferred Value Preferred
case preference order value in plan preference order value in plan
1 v1 > v4 > v3 > v2 v1 inP L1 v3> v4> v1 > v2 v3 inP L4
2 v3 > v2 > v1 > v4 v3 inP L2 v3> v4> v1 > v2 v3 inP L4
3 v2 > v1 > v3 > v4 v2 inP L1 v2> v4> v1 > v3 v2 inP L5
4 v3 > v1 > v4 > v2 v3 inP L2 v1> v2> v4 > v3 v1 inP L4
5 v1 > v4 > v3 > v2 v1 inP L1 v2> v4> v3 > v1 v3 inP L4
From the list critical questions in Section 3.5, twenty eight questions were imple- mented for these experiments. I implemented critical questions related to validity, pos- sibility and suitability for the action and plan specifications, alternative plans and side effect questions. Critical questions regarding action-combination and time-related ques- tions to avoid the complexity of implementing those characteristics for the example.