Fast and Learnable Behavioural and Cognitive Modelling for Virtual character Animation

1
Fast and Learnable Behavioural and Cognitive
Modelling for Virtual character Animation
Distributed Computational Intelligence and
Knowledge Discovery
2

Distributed Computational Intelligence and
Knowledge Discovery
AGENDA

• Introduction

• Cognitive Modelling (CM)

• Fast CM Approximation with Artificial Neural
  Networks (ANN)

• Character Learning

3

Distributed Computational Intelligence and
Knowledge Discovery
AGENDA

• Introduction

• Cognitive Modelling (CM)

• Fast CM Approximation with Artificial Neural
  Networks (ANN)

• Character Learning

4

Distributed Computational Intelligence and
Knowledge Discovery
Virtual Characters
5

Distributed Computational Intelligence and
Knowledge Discovery
Virtual Characters
6

Distributed Computational Intelligence and
Knowledge Discovery
Virtual Characters
Model name Laelia Geographic area Europe
Height 173cm Hair colour Chestnut Kind
of hair Wavy Hair lenght Long Eye
Colour Green-Brown Breast 93 Waist
62 Hips 92 Distinguishing mark Modesty,
open-mindedness Hobby Modern art Status of
inscription In Competition
7
Distributed Computational Intelligence and
Knowledge Discovery

• Challenges of Animating Virtual Characters
• Very demanding and time consuming to animate
  complex models
• Need for autonomous intelligent behaviour in a
  dynamic environment
• Current behavioural and Cognitive modelling have
  serious limitations
• Use Computational Intelligence methods to
  overcome these limitations

8

Distributed Computational Intelligence and
Knowledge Discovery
AGENDA

• Introduction

• Cognitive Modelling (CM)

• Fast CM Approximation with Artificial Neural
  Networks (ANN)

• Character Learning

9
Distributed Computational Intelligence and
Knowledge Discovery
Cognitive Modelling

• Strategic Reasoning and planning in contrast to
  the tactical reactionary approach of behavioural
  models.
• Defines what a character knows, how the knowledge
  is acquired and used to plan actions.
• Planning formulates optimal sequence of actions
  to desired goal i.e. target state

10
Distributed Computational Intelligence and
Knowledge Discovery
Cognitive Modelling

• Set of branching actions must be kept very small
  for real time performance
• Performance trade-off leads to suboptimal plans
• Limits on number of intelligent characters in the
  animation
• Inability to learn new models autonomously

11

Distributed Computational Intelligence and
Knowledge Discovery
AGENDA

• Introduction

• Cognitive Modelling (CM)

• Fast CM Approximation with Artificial Neural
  Networks (ANN)

• Character Learning

12
Distributed Computational Intelligence and
Knowledge Discovery
Artificial Neural Networks
Natural Neuron
Artificial Neuron
13
Distributed Computational Intelligence and
Knowledge Discovery
Artificial Neural Networks
3 Layer Feedforward Artificial Neural Network
14
Distributed Computational Intelligence and
Knowledge Discovery
Fast Character Animation of CM with ANN

• CM with goal defines policy

a µ(I) a action to perform I current
state

• Feed forward neural network is used to
  approximate µ(I)

• Selection of appropriate inputs requires domain
  knowledge

15
Distributed Computational Intelligence and
Knowledge Discovery
Fast Character Animation of CM with ANN
After training the neural network is
incorporated into the brain of the virtual
character
16
Distributed Computational Intelligence and
Knowledge Discovery
Fast Character Animation of CM with ANN
To approximate different policies different
neural networks are used and incorporated into
the brain of the virtual character
17

Distributed Computational Intelligence and
Knowledge Discovery
AGENDA

• Introduction

• Cognitive Modelling (CM)

• Fast CM Approximation with Artificial Neural
  Networks (ANN)

• Character Learning

18
Distributed Computational Intelligence and
Knowledge Discovery
Offline Character Learning

• A character automatically learns an unknown
  cognitive or behavioural model
• Very important because it can reduce the amount
  of work and time spent to develop an animation
• Done using reinforcement learning and artificial
  neural networks for generalization

19
Distributed Computational Intelligence and
Knowledge Discovery
Reinforcement Learning

• Agent (virtual character in this case)
  automatically approximates a fitness function by
  maximising its cumulative rewards
• Fitness function determines new cognitive model
• Methods used include Q learning, and tree search
  algorithms

20
Distributed Computational Intelligence and
Knowledge Discovery
Character Learning Demo
21
Distributed Computational Intelligence and
Knowledge Discovery
Literature and Resources

• Fast and learnable behavioral and cognitive
  modeling for virtual character animation
  Jonathan Dinerstein, Parris K. Egbert, Hugo de
  Garis, Nelson Dinerstein Computer Animation and
  Virtual Worlds 2004 15 95108
• Spaceships Animations http//neptune.cs.byu.edu/e
  gbert/a3dg/projects/bcm/animation.html
• Neural networks http//tracer.lcc.uma.es/problems
  /ann/ann.html http//www.geocomputation.org/2000/
  GC016/Gc016.htm
• Reinforcement Learning http//www.cs.ualberta.ca/
  sutton/book/the-book.html
• Miss digital world virtual characters and
  animation http//www.missdigitalworld.com/