CS7637 KBAI Spring 2018: Project 2 Reflection
1
John Tan
Professor Ashok Goel
CS 7637: Knowledge-Based AI
March 18, 2018
Project 2 Reflection
Project 1 Overview
In project 1, I approached the Raven Progressive Matrice (RPM) using the visual
approach. The visual approach generates a 2x2 agent that creates both vertical and
horizontal relationship from A to C and A to B. It use the generate and test approach,
along with the affine method to solve RPM, where the problem figures transform,
rotate, and reflect from A to B or from A to C that compares to the solution figure of
1 to 6, which I call it “D” in the diagram using the root mean square (RMS). The
following vertical and horizontal relationships that the 2x2 agent generates are as
follows:

CS7637 KBAI Spring 2018: Project 2 Reflection
2

CS7637 KBAI Spring 2018: Project 2 Reflection
3
Throughout the image comparisons of AB with CD or AC with BD, it uses the RMS
equation. The RMS equation that use to measure the difference between two images
is define as the following:
Homes, S., 2000, Stanford
The RMS error is always between 0 and 1. If AB and CD (#) or AC and BD (#) has a
RMS close to 0, it means that the comparison of AB and CD (#) or AC and BD (#) is
closely related to the solution ‘#’ that is chosen as figure ‘D’.
Changes in Problem Set & Agent’s Reasoning
The complexity level from Project 1 to Project 2 increases from a 2x2 to a 3x3
problem set, which adds another layer of complexity and relationships. Instead of
only vertical and horizontal, it is now include a diagonal direction in the 3x3 grid.
The number of choices also increases from 6 to 8. In Project 2, I continue to use the

CS7637 KBAI Spring 2018: Project 2 Reflection
4
visual approach using a different knowledge of representation to solve RPM. Instead
of using generate and test from the previous 2x2 problems, I use a similar approach
of analysis, which is the Gestalt method where each representations of the figures
are visually compared. This visual strategy for RPM “uses visual abstractions over
problems to approximate the answer even without precise knowledge of the
transformation between frames” (Joyner, D., 2015, GA Tech). This strategy mirrors
the human cognition where we look at the trends of likelihood of the correct answer
by taking a list of measurements for each individual figure and compares it against
the solution figures. If the answer returns the minimum, it is the chosen answer. The
measurement I use for each individual problem figures and solutions to compare is
the Dark Pixel Ratio. Since I didn’t take individual shapes and edges into accounts
and only care about the dark pixels, it made it easier to compute the likelihood of
the possible answer when comparing the solutions with the “base” figures of A, B, C,
D, E, and F. The idea of the “Dark Pixel Ratio” (DPR) comes form Dr. David Joyner’s
research paper “
Using Human Computation to Acquire Novel Methods for Addressing
Visual Analogy Problems on Intelligence Tests
”, which describes as “the difference in