Evaluating and re-evaluating agent modeling: simulation and design - PowerPoint PPT Presentation

Loading...

PPT – Evaluating and re-evaluating agent modeling: simulation and design PowerPoint presentation | free to download - id: 595627-YTI4Y



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Evaluating and re-evaluating agent modeling: simulation and design

Description:

Title: Evaluating and re-evaluating agent modeling: (or) the impact of cognitive models on generative architectural design Author: Dan B Last modified by – PowerPoint PPT presentation

Number of Views:568
Avg rating:3.0/5.0
Slides: 38
Provided by: DanB49
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Evaluating and re-evaluating agent modeling: simulation and design


1
Evaluating and re-evaluating agent modeling
simulation and design
Daniel Belcher
?
January 11th, 2007 Arch 484 Design Computing
Seminar
2
Simulations and ModelsVisual Models
Ballard Library Neighborhood Service Center.
Bohlin Cywinski Jackson, Architects
3
(No Transcript)
4
UW project http//www.urbansim.org/
5
ECOTECT.com
Physical simulation
6
Old-school Mechanisms
  • For seeing life is but a motion of limbs, the
    beginning whereof is in some principal part
    within, why may we not say that all automata
    (engines that move themselves by springs and
    wheels as doth a watch) have an artificial
    life?"
  • -Thomas Hobbes, Leviathan, 1660.
  • Lhomme est une machine.
  • Man is a machine.
  • -Julien Offray de La Mettrie, Lhomme machine,
    1748.
  • Verum et factum convertuntur.
  • The true and the made are convertible.
  • -Giambattista Vico, De nostri temporis studiorum
    ratione, 1709.

7
(No Transcript)
8
MASSIVE Software
Agents on film
Fellowship of the Ring battle scenes by Weta
digital.
9
Its all soooooo pedestrian
  • Much of agent modeling is focused on navigation,
    locomotion, and movement through space.
  • Why? Humans are extremely complexand even
    walking around is difficult to model.
  • However
  • The dynamics of pedestrian crowds are
    surprisingly predictable

10
  • Pedestrian activity can be modeled as a
    self-organizing system (Helbing et al, 2001).

Time-lapse photography standing crowd outside a
movie theater showing crossing pedestrians
forming a river-like flow.
Agent models of pedestrian flows.
11
EVAS Pedestrian Modeling Software
http//www.vr.ucl.ac.uk/research/evas/evas.html
12
  • What behavior should the agent simulate?
  • Does the agent exhibit this behavior?
  • Do humans behave in the same way?
  • How do groups of humans behave?
  • Do models exhibit these group behaviors?
  • Can models capture something beyond simply
    behavior?
  • Can they capture emotion? Mood? Cognitive
    process?
  • Just how predictable are people?
  • Should we model agents at all?
  • What assumptions does agent modeling make?

13
Two types of control
Model-based control
Planned path
Emergent path
14
Sense-Model-Plan-Act
Model
environment
agent
Adapted from (Russell Norvig, Artificial
Intelligence,1995)
15
The Ecological approachS(P)A
???
AKA Gibsons direct perception (Gibson, The
Ecological Approach to Visual Perception,
1979) AKA Active Perception in
robotics (Brooks, Intelligence without
representation, 1991) Subsumption
architecture AKA Situated, reactive agents
16
Behaviors as rules
(Reynolds, Flocks, herds, and schools A
distributed behavioral model, 1987)
a) Separation. Steer to avoid local
flock-mates. b) Cohesion. Steer to move toward
the average position of local flock-mates. c)
Alignment. Steer toward the average heading of
local flock-mates. d) Avoidance. Steer to avoid
running into local obstacles or non-flock-mates.
1) Pedestrians are motivated to move as
efficiently as possible to a destination. 2)
Pedestrians wish to maintain a comfortable
distance from other pedestrians. 3) Pedestrians
wish to maintain a comfortable distance from
obstacles. 4) Pedestrians may be attracted to
other pedestrians or objects.
17
Evaluating agent modeling
18
Why does all this matter?
Answer Agent-based simulation allows designers
to evaluate the behavior of individuals and
groups inhabiting a space.
  • Learn more about the agent-environment dynamic
  • Validate new designs against known behavior from
    old designs
  • Better understand and improve upon existing
    buildings
  • Help train building operators to better manage
    their buildings
  • Generate building visualizations showing
    life-like usage patterns
  • Illustrate the consequences of changes to
    building structure

19
MouseHaus
(Therakomen, 2000 2001).
Pros Seeks to model reflexive, reactive and
motivated behaviors. Computationally
efficient. - Cons Agent steering dynamics are
simplistic. Linear behaviorno
learning.
20
Agent-based Virtual Users
(Yan and Kalay, Simulating Human Behavior in
Built Environments, 2005)
21
Artificial Life Behavior Modeling primary
movement control was flocking (as in Reynolds,
1987). B f(G,R,E)
International standard of human
modeling Humanoid Animation Specification
(H-Anim, 1.1)
22
3D visual simulation of plaza, with and without
fountain. (Yan and Kalay, 2005).
  • Pros
  • Interactive simulation.
  • Uses standard media (DXF) BIM.
  • Conducted study of observed behavior.
  • Cons
  • Artificial life model is extremely simplistic.
  • Agents explore, but do not learn.
  • Affordances are explicitly encoded in the
    environment, and not as emergent behavior.

23
Curious Agents
(Saunders and Gero, Curious Agents and Situated
Design Evaluation, 2004)
  • Exploratory agents
  • Ray-casting perception
  • Curiosity model (Saunders and Gero, 2001)
  • Learning model
  • Exploring an art gallery

24
Agent Evaluations, before and after
  • Agents Post-Occupancy Evaluation
  • Even dispersal of interest
  • Less crowding
  • All rooms visited by each agent
  • Agents learn a random array of art work
  • Uneven dispersal
  • Crowding around entrance and exit
  • Stuck in local-minima (NW room empty)

25
Ecoconfiguration Generative Design
(Turner, Mottram Penn, An Ecological Approach
to Generative Design, IJDC, 2004)
  • Generative Component
  • Environments are randomly seeded.
  • Genetic Algorithm employed to optimize
  • configurations.
  • Spatial syntax used as fitness function
  • Axial arrangements selected for.
  • (Penn Turner, 2002)

Simple Agent Affordances walkable and
seeable Walk three steps, look around, repeat
26
Foyer?
(from Turner, Mottram and Penn, 2004.)
27
Re-evaluating agent modeling
28
Two types of controlrevisited
Model-based control
Planned path
Emergent path
29
Behavioral Dynamics
  • Behavioral variables (Schöner, Dose Engles,
    1995)
  • goals expressed as (sets of) points in space
    spanned by behavioral variables
  • behavior corresponds to trajectories through that
    space
  • Behavioral Dynamics
  • (Fajen and Warren, 2001)
  • trajectories expressed as solutions to system of
    differential equations
  • attractors (intended states) and repellors
    (avoided states)
  • behavior emerges as a consequence of how
    information is used to adjust action system

30
Behavior corresponds to trajectories through that
space?
31
Trajectories expressed as solutions to system of
differential equations?
Mass-spring-damper
Sinusoid oscillation
32
Dynamics of Steering Obstacle Avoidance
  • Behavioral variables
  • Heading (f) and its rate of change (?)

Fixed exocentric frame of reference
.
heading
f
  • Behavioral dynamics
  • Identify factors
  • Develop an equation of motion
  • Predict routes
  • (Fajen Warren, 2001)

Dynamics of Steering
33
The Steering Model (Fajen Warren, 2001)
34
The VENLab
InterSense 900 Tracker sonic beacons (12 x 12 m)
microphones inertia cube
Kaiser Proview 80 HMD stereo (60 x 40)
  • Manipulate goals obstacles during walking
  • Record paths x and z position data

35
Random Obstacle Fields (Warren Belcher, 2002)
8
6
4
Z (m)
2
0
-
-
4
2
0
2
4
X (m)
  • 8 random arrays, forward and backward
  • 15 trials per condition, Number of subject 10
  • 8 15 120 trials per subject10 subjects
    1200 reps
  • Our goal Observe and predict paths

36
Random Obstacles Findings
S8
Z (m)
Array 1
-2
-1
0
1
2
-2
-1
0
1
Z (m)
Array 2
-1
0
1
2
-1
0
1
2
X (m)
X (m)
individual differences ? different set of
parameters
65 of all human paths are within 1 obstacle of
model
37
Why does all this matter again?
Agent-based simulation allows designers to
evaluate the behavior of individuals and groups
inhabiting a space.
Important to iteratively re-evaluate agent
modeling on the basis of emergent models from
cognitive science and robotics.
Deepens our understanding of the dynamic
complexity of human activity and our coupling
with the Built Environment.
38
Thank
You
39
Goal Experiment (Fajen Warren, 2001)
goal
heading
dg 2, 4, 8 m
?g 5, 10, 15, 20, 25
  • Observers begin walking in a given direction
  • After a few steps, a goal appears
  • Vary initial goal angle and distance
  • Instructions Walk to goal.

40
Goal Experiment - Findings
8
5
25
8 m
4
6
4 m
3
z (m)
4
z (m)
2 m
2
2
1
4 m condition
20 condition
0
0
0
1
2
3
0
1
2

x (m)
x (m)
  • Goals function as attractors of f
  • Acceleration toward goal increases with goal
    angle
  • Acceleration toward goal decreases with goal
    distance

41
Single Obstacle (Fajen Warren, 2001)
goal
obstacle
9 m
do 3, 4, 5 m
?o 1, 2, 4, 8
  • Begin walking toward goal
  • After a few steps, an obstacle appears
  • Vary initial obstacle angle and distance
  • Instructions Walk to goal while avoiding
    obstacle.

42
Single Obstacle Experiment Findings
  • Obstacles function as repellors of f
  • Acceleration away from obstacle decreases with
    obstacle angle
  • Acceleration away from obstacle decreases with
    obstacle distance

43
Two Obstacles (Fajen Warren, 2001)
goal
obstacle 2
obstacle 1
Model predictions
8
small angle
large angle
medium angle
6
middle
left
right
4
2
0
44
Two Obstacles Findings
0
2
4
50
37
0
63
0
50
69
2
29
8
6
10
65
19
16
48
46
6
34
64
1
2
-2
-1
0
1
Humans switch from right ? left ? center route
as obstacle angle increases
About PowerShow.com