1. 컴퓨터 그래픽스 분야의 캐릭터 자동생성을 위하여 인공생명의 여러 가지 방법론이 어떻게 적용될 수 있는지 이해
제5주. Art and Design An Artificial Life Approach for the Animation of Cognitive Characters
F.R. Miranda, J.E. Kogler Jr, E.D.M. Hernandez and M.L. Netto, Computers & Graphics, vol. 25, pp. 955~964, 2001
학습목표
컴퓨터 그래픽스 분야의 캐릭터 자동생성을 위하여 인공생명의 여러 가지 방법론이 어떻게 적용될 수 있는지 이해

2. 개요 Cognitive character animation Perception : neural networks
Behavioral control : finite state machine optimized by genetic algorithm
WOXBOT/ARENA research project
Build virtual worlds where small robots perform tasks with their own motivation and reasoning
Open distributed object architecture for
Evolutionary computation, artificial life, pattern recognition, artificial intelligence, cognitive neuroscience, distributed objects architecture

3. Introduction
Goal of artificial life: Computational model of complex behavior
Simulation of macroscopic behavioral aspects of living beings using microscopically simple components
Virtual characters whose behavior emerges from hierarchical and functionally specialized complex structures
Artificial life worlds
Virtual places where animated characters interact with the environment and with other virtual beings of the same or distinct categories
Behavioral animation: characters have some degree of autonomy to decide their actions
Cognitive animation
AI + evolutionary computation + graphics
Perform tasks not explicitly specified
Adaptive algorithms + evolutionary programming + AI

4. Introduction (2) ARENA : Fig. 1
Artificial environment for animated virtual characters
Goal: obtain virtual creatures capable of performing specified tasks in their environment by exploring certain strategies and adapting them
WOXBOT : Fig. 2 (Wide Open eXtensible roBOT)
Vision system: simulated camera and neural network
Motor system: finite state machine optimized by GA
Motion: forward, backward, turn left, turn right
Purposes
Behavior modeling: lab of learning algorithms
Research on societies of virtual characters
Study of collective dynamics of populations

5. Artificial Life A-life in computer science
Cellular automata theory  computer graphics animation
Universal life concept  evolution and natural selection concepts
GA with mutation and combination to change one generation to another
Keeping greater energy, living more time, performing tasks faster, …
Use of knowledge  intelligent behavior
Present in environment and in conception of creatures
Used by creatures when performing actions

6. Intelligent Agents Definition
Computational entities of autonomous behavior in agent design, environment, goals and motivations
Life features: able to sense world, analyze the information, able to express the decisions through actions
Sensing and perception
Vision and audition  neural networks (3-layer NN) to recognize pyramids and cubes
Sensing  rendering better images
Perception  adjusting NN size and training
Behavior: reasoning and acting by learning and evolution
Learning and evolution through generations
Natural selection

7. Project Overview Requirements Mathematical and computational models
Efficient environment and platform
Constant evolution and improvement
ARENA implementation
Distributed communicating objects on microcomputer cluster
Parallelism through multithreading
Floor and walls, objects (obstacles, barriers, traps, shelters, energy or food sources, …)
WOXBOT: task is to keep itself alive as long as it can
Major ingredients
Sensing and perception, knowledge use and evolution

8. Current Implementation Model
Sensing and perception (pattern recognition)
Nutrients (yellow pyramids) and hurting entities (red cubes)
JAVA3D  3 color channels of RGB: Fig. 3
2 specialized neural networks for 4 outputs: Table 1
ANN-I: targeting nutrients (identification of yellow pyramids)
ANN-II: targeting hurting entities (for red cubes)
1 hidden layer with 8 nodes
Use of samples of similar sizes: no scale invariant
Action and behavior: learning and reasoning
Finite state machine
Input: combination of ANN-I and ANN-II (Table 2)
Output: 00 (turn left), 01(go straight), 10 (turn right), 11 (go back)
Evolving FSM from initial random structure with life duration
Chromosome: entry for 16 possible inputs (next state, action code)
Fig. 4 & 5

9. Implementation Issues
Real-time application?
Maya, Softimage  good API  bad user interaction
Choice of API
Portability, share ability, open architecture, easy of use
JAVA PL, JAVA3D for graphics
Scene-graph oriented approach  shortening development time
Environment: ARENA
Green rectangular floor-plan, four blue walls, red cubic boxes, yellow pyramids
Looking for pyramids while avoiding cubes
Character: WOXBOT
FSM optimized by GA
States, inputs and actions

10. Simulation Results and Further Work
GA parameters
16 individuals, 30 generations
Half: random generation, half: crossover and mutation (0.06)
Fig. 7 & 8
Fig. 9: JAVA3D to develop ARENA/WOXBOT project
Limitations
No memory of the previous actions  no learning
Low number of states  only 4
Size variant
Distribution of objects
Textures
Multiple WOXBOTs