Project Description

This is the repository of the group project of the "Intelligents Agents" course at Utrecht University, academic year 19/20. The project consists in designing and implementing a personal assistant agent that will help a new master student choose courses. Our solution is an OWL 2 Web Ontology implemented within the Protégé framework of Stanford university with the Manchester syntax. The user interface has been implemented in Java.

User Guide

The UI is as below:

• Query can be performed by checking any preferences in the checkbox.
• Register can be performed by clicking the "Register" button when one or multiple courses have been checked.
• "Ask for Evaluation" can be performed only in the multi-agent environment.

To run this application in the one-agent environment:

1. Run the project course-planner per README.md inside
2. Run the project agent-discovery per README.md inside
3. Open and browser and go to this link: http://localhost:8080

Agent Design

Agent Architecture

Rational agents come into the following five types:

• Reflex agents
• Reflex agents with state
• Goal-based agents
• Utility-based agents
• BDI agents

The course planner is built upon the architecture of utility-based agents.

The state consists of agent state and world state, which is able to be perceived by the sensors from the environment:

• world state
  • S1: Online neighbouring agents
  • S2: Course evaluation from neighbouring agents
• agent state
  • S3: Total registered courses
  • S4: Available time slots in the calendar
  • S5: Reputation rating of neighbouring agents

The action stands for the action that the agent is able to take, which will influence the internal agent states and the environment by the actuators:

• A1: Register for a course
• A2: Ask for the evaluation of a course from neighbouring agents
• A3: Answer the evaluation of a course from neighbouring agents
• A4: Give positive or negative feedback for evaluations of courses from neighbouring agents

How the world evolves (which is perceived by the sensors) and what my actions do (which is actuated by the actuators) will both have an effect on the world and agent states:

Sensors	Actuators	State
E1: Neighbouring agents online		S1: Online neighbouring agents
	A2: Ask for the evaluation of a course from neighbouring agents A3: Answer the evaluation of a course from neighbouring agents	S2: Course evaluation from neighbouring agents
	A1: Register for a course	S3: Total registered courses S4: Available time slots in the calendar
	A4: Give positive or negative feedback for evaluations of courses from neighbouring agents	S5: Reputation rating of neighbouring agents

The utility is calculated as below:

• Utility = 0.8*(number of requirements respected / total number of requirements) + 0.2*(sum(1 if there is at least a friend in course i, 0 otherwise) / number of courses)

The knowledge base is imported from the ontologies, which contains the following consistent facts about a student and courses, which will be queried by the agent and inference will be drawn based on the chosen reasoner: HermiT, Pellet or Fact++:

• Facts about a student:
  • Preferences for a certain lecture, day and topic
  • Friends
• Facts about a course:
  • Lecturers
  • Faculties
  • Research Methodologies
  • Topics
  • Time slots

Computational trust comes into the following three types:

• Local trust
• Institutional trust
• Social trust

Beta-reputation system is based on social trust:

• Beta density function is used for binary events
• Count the past occurrences and decide based on frequency

Formula to calculate reputation and opinion is as below:

• r = occurrences of positive feedback, s = occurrences of negative feedback
• α = r + 1, β = s + 1

Y's reputation calculated by X:

• reputation expected: E(p) = α / α + β, the result is between [0, 1]
• reputation rating: Rep = (E(p) - 0.5) * 2, the result is between [-1, 1]

X's opinion about Y for belief, disbelief, uncertainty: w = (b, d, u)

• b = r / (r + s + 2)
• d = s / (r + s + 2)
• u = u / (r + s + 2)
• mapping from opinion to reputation: Rep(r, s) = (2b/u, 2d/u)

Combining opinion: X's opinion about Y = (bxy, dxy, uxy), Y's opinion about Z = (byz, dyz, uyz), and X's opinion about Z = (bxz, dxz, uxz):

• bxz = bxy * byz
• dxz = bxy * dyz
• uxz = dxy + uxy + bxy * uyz

Forgetting factor based on timestamp i is as below:

• r = r1 + ... + rn, s = s1 + ... + sn
• r = (r1 + ... + rn) ∗ λ⁽ⁿ⁻ⁱ⁾, s = (s1 + ... + sn) ∗ λ⁽ⁿ⁻ⁱ⁾, where 1 ≤ i ≤ n, 0 ≤ λ ≤ 1

Rating-based approaches:

• Sensitive to subjectivity
• Sensitive to variation in context

Ontology-based experiences:

• Gaussian Model: It estimates the probability of satisfaction
• Case-Based Reasoning (CBR): Score for experience = recency * similarity * satisfaction

Scenarios

[to be completed...]

Evaluation

• Correctness ratio

Performance Metrics

[to be completed...]

References

More information about the Manchester syntax and the OWL 2 Web Ontology language can be found in the following links. http://ceur-ws.org/Vol-216/submission_9.pdf https://www.w3.org/TR/owl2-manchester-syntax/

• https://owlcs.github.io/owlapi/apidocs_5/index.html