# Matt Lehman.pdf - KBAI Assignment 1 Solving Verbal and/or...

• Essay
• 4
• 100% (2) 2 out of 2 people found this document helpful

This preview shows page 1 - 3 out of 4 pages.

KBAI - Assignment 1Solving Verbal and/or Visual Representations of 2x2RPM using Generate and Test with SemanticNetworksMatthew LehmanAbstractA design for an agent capable of solving 2x2 Raven’s Progressive Matrices (RPM). The agent uses geometricanalogy networks, a specialized type of semantic network, to represent patterns of relationships in and transfor-mations between images. In order to find the best solution, the agent uses generate and test to produce all validpatterns as geometric analogy networks for the examples and solutions using the choices. The agent then teststhe similarity of example and solution networks to find the best match.ContentsIntroduction11Semantic Networks21.1Challenges. . . . . . . . . . . . . . . . . . . . . . . . . . .21.2Geometric Analogy Network. . . . . . . . . . . . . . .21.3Implementation. . . . . . . . . . . . . . . . . . . . . . . .22Generate and Test32.1Challenges. . . . . . . . . . . . . . . . . . . . . . . . . . .32.2Process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32.3Implementation. . . . . . . . . . . . . . . . . . . . . . . .33Other Considerations43.1Verbal Explanations of Match. . . . . . . . . . . . . .43.2Limits of Geometric Analogy Networks. . . . . . .43.3Similarity of Objects when Matching. . . . . . . . .43.4Learning from Mistakes. . . . . . . . . . . . . . . . . .43.5Learning Transformations. . . . . . . . . . . . . . . . .4References4IntroductionThis is an overview of a design for an agent that is capableof solving 2x2 problems from Raven’s Progressive Matrices(RPM). The problems are visual analogies (see Figure1)where four images are arranged in a 2x2 grid, where eachimage contains an arrangement of objects of varying shapesand other properties (e.g.fill, size).The grid of imagesare related to each other both horizontally and vertically byan unknown set of geometric transformations, including theaddition, removal, movement, scaling, reflection, and rotationof the objects. The agent is given examples containing the firstthree images (A,B and C) and must then select the best choicefor D from a set of images. The best choice is the image thatcompletes a pattern of horizontal and vertical transformationsin the simplest and most consistent way.Figure 1.A Visual Analogy ProblemABC?D123In order to accomplish the task, the agent must be ableto complete two high level steps (see Algorithm1). First
Solving Verbal and/or Visual Representations of 2x2 RPM using Generate and Test with Semantic Networks — 2/4it must be able to take the raw images, extract necessaryfeatures, and build a representation of problem. Second it mustapply a problem solving method to the representation to derivea solution. To simplify the challenge of feature extractionfrom raw images, Verbal representations are available as inputwhere the challenging task of visual feature extraction hasalready been performed by a human. The rest of this paperwill assume the use of the verbal representations.Algorithm 1:Solving Visual Analogy ProblemsInput: Visual representation of problemOutput: Best choiceextract object features from images;generate representations from features;test representations for match;returnimage producing best match;