This document proposes applying a computational model of consciousness called A-Consciousness to design more human-like and engaging game characters. It describes an architecture called CERA-CRANIUM that implements A-Consciousness and applies it to control game characters in Unreal Tournament 2004. An evaluation found that characters controlled by CERA-CRANIUM performed comparably to human players, outperforming other AI control methods. The authors conclude the approach is promising but needs to incorporate more advanced learning and memory capabilities.

1. Towards Conscious-like
Behavior in Computer Game
Characters
Raúl Arrabales, Agapito Ledezma and Araceli Sanchis
g p
Computer Science Department
Carlos III University of Madrid
http://Conscious-Robots.com/Raul
htt //C i R b t /R l
IEEE Symposium on Computational Intelligence and Games 2009
Politecnico di Milano, Milano, 9th September 2009

2. Contents
Introduction
Motivation and Objectives
State of the Art
Consciousness and Games
Experimental Environment
Proposed Architecture
Evaluation
Conclusions
www.Conscious-Robots.com 2

3. Introduction (I)
Objective:
Design more appealing synthetic characters.
mo e s nthetic cha acte s
More engaging, more real.
Human-like (Conscious-like behaviors).
H lik (C i lik b h i )
Problem:
Many cognitive s s involved.
a y cog t e skills o ed
How to integrate them effectively?
www.Conscious-Robots.com 3

4. Introduction (II)
Character design challenges:
Emotions, planning, learning, opponent
E ti l i l i t
modeling, attention, set shifting, etc.
How can they be integrated effectively?
Our hypothesis:
Use a cognitive architecture based on a
model of consciousness.
www.Conscious-Robots.com 4

5. Motivation and Objectives
Ultimate goal:
Characters
Cha acte s able to pass the (adapted) Turing test.
T ing test
Why?
State of the art video games
g
Complex virtual environments.
Elaborated scripts.
Require more realistic behaviors.
www.Conscious-Robots.com 5

6. State of the art
Current AI characters cannot match human-
like behavior
behavior.
E.g. see last year 2K Botprize contest results.
Usually, playing with other humans is
more realistic and engaging that playing
with AI bots.
Apply Machine Consciousness to games to
improve believability?
www.Conscious-Robots.com 6

7. Consciousness and Games (I)
Embodiment and situatedness are claimed to be
critical for the production of consciousness.
p
We argue that consciousness based software
consciousness-based
agents can also be embodied and situated.
Software sensors and actuators
actuators.
Causal interaction with the world.
Virtual World (virtually situated).
Real World (interaction with human players).
www.Conscious-Robots.com 7

8. Consciousness and Games (II)
Games versus Physical Robots
Robots possess physical body.
body
Robots interact directly with physical world.
Robots have to deal with noise and uncertainty.
uncertainty
These aspect can b seen as benefits or
Th t be b fit
drawbacks
Games are ideal for focusing on h h l
d lf f high-level control.
l l
www.Conscious-Robots.com 8

9. Experimental Environment
FPS: Unreal Tournament 2004
Bot Behavior Controller
Basic Behaviors
B i B h i
Sensory Data
Processing Movement Animation
Module Module
Sensory Body
Inputs State Atomic Actions
Virtual World
www.Conscious-Robots.com 9

10. Proposed Architecture (I)
Context: Machine Consciousness Research.
Specifically: Design of machines showing conscious-like
p y g g
behavior.
Access Consciousness (A C
A C i A-Consciousness) Approach:
i A h
Only a small selection of contents gain access to a unique
sequential thread for explicit processing and volition.
q p p g
Computational model based on:
Global Workspace Theory (Baars, 1988).
Multiple Draft Model (Dennett, 1991).
www.Conscious-Robots.com

11. Proposed Architecture (II)
Computational model of consciousness:
Large number of specialized processors.
L b f i li d
Competing and collaborating to access a
g
global workspace (working memory).
p ( g y)
Where coherent information patterns are
selected.
www.Conscious-Robots.com

12. Proposed Architecture (III)
Contexts
Working Memory
Focus of
Attention
Specialized
Processors
Interim
Coalition
www.Conscious-Robots.com

13. Proposed Architecture (IV)
CERA-CRANIUM Architecture
Implements the computational model of A-
A
Consciousness.
CERA is a layered control architecture.
CRANIUM is a runtime component for the
management of specialized processors.
g p p
www.Conscious-Robots.com

14. Proposed Architecture (V)
CERA-CRANIUM Perception Flow
Agent’s
CERA World
Body
B d
Physical Layer Sensor
Sensors
S Accessible
Mission‐specific Services environment
Motor Reachable
Core Layer Actuators environment
Services
www.Conscious-Robots.com

15. Proposed Architecture (VI)
CRANIUM Workspace (at CERA Physical layer)
CRANIUM
Single
Single Workspace Complex
Complex
Percepts Percepts
Percept
… Aggregators
www.Conscious-Robots.com

16. Proposed Architecture (VII)
CRANIUM Workspace Modulation
Modulation Commands
Core
Workspace Workspace Layer
Single Complex Mission
Percepts Percepts Percepts
… …
Physical Layer Mission‐Specific Layer
www.Conscious-Robots.com

17. Proposed Architecture (VIII)
CRANIUM Workspace Modulation
UT2004 Server
CERA
UT2004 Viewer
UT2004 Viewer GameBots
2004
CERA‐CRANIUM
CERA CRANIUM UT2004 Client
UT2004 Client
CCR / DSS Translated (Human player)
Runtime Pogamut 2
UT2004 Client
UT2004 Client
.NET Framework
(Spectator)
www.Conscious-Robots.com

18. Proposed Architecture (IX)
CERA-CRANIUM instantiation for UT2004
CERA sensory-motor services layer.
CERA physical layer specialized processors.
CERA mission-specific layer specialized
p y p
processors.
CERA core layer goals and model state
state.
www.Conscious-Robots.com

19. Proposed Architecture (X)
Example of percepts for UT2004
SP(Being Damaged) +
SP(S E
See Enemy) +
CP(Enemy Approaching) =
MP(EEnemy Att ki )
Attacking
www.Conscious-Robots.com

20. Evaluation (I)
Assessing believability is not straighforward.
Subjective process
process.
Depends on opponents behavior.
Game score is not significant.
AVERAGE
PLAYER
SCORE
Rule-Based System Bot 19.2
Q-Learning Bot 5.2
Q-Learning and Expert Systems Hybrid Bot 11.3
CERA-CRANIUM
CERA CRANIUM Bot 18.3
18 3
www.Conscious-Robots.com

21. Conclusions
The application of consciousness-inspired
cognitive architectures is promising
promising.
We
W need to incorporate learning algorithms
d i l i l ih
and long term (episodic and semantic)
memory t th model.
to the d l
Improve ToM (opponent/friend modeling).
www.Conscious-Robots.com

22. Thank you
www.Conscious-Robots.com 22