new bulgarian university mindraces, first review meeting, lund, 11/01/2006 anticipation by analogy...
TRANSCRIPT
MindRACES, First Review Meeting, Lund, 11/01/2006
New Bulgarian University
Anticipation by AnalogyAnticipation by Analogy
An Attempt to Integrate Analogical Reasoning with Perception, Selective Attention, Context, and Motor Control
An Attempt to Integrate Analogical Reasoning with Perception, Selective Attention, Context, and Motor Control
MindRACES, First Review Meeting, Lund, 11/01/2006 2NBU
Anticipation Mechanisms
• Explicit Anticipation: analogy-making Predictions based on one single example• Implicit Anticipation: context & relevance Predicting relevance based on context –
guiding attention in reasoning and perception• Combining Explicit and Implicit
Anticipation
MindRACES, First Review Meeting, Lund, 11/01/2006 3NBU
Examples of Anticipation based on analogy-making and context
• Searching for your keys
They are not at their usual place, so • try to reconstruct what you have done with
them (memory reconstruction), • reminding of old episodes of key search
and where you found them (analogy)• Perceived elements (context) guide the
reconstruction, reminding, and analogy
MindRACES, First Review Meeting, Lund, 11/01/2006 4NBU
Examples of Anticipation based on analogy-making and context
• Searching for your car in the parking slot
• try to reconstruct where you have parked it (memory reconstruction),
• reminding of old episodes of car search and where you found it (analogy)
• reminding of old episodes of key search and where you found it (remote analogy)
• Perceived elements (context) guide the reconstruction, reminding, and analogy
MindRACES, First Review Meeting, Lund, 11/01/2006 5NBU
Examples of Anticipation based on analogy-making and context
• Predicting the outcome of a game
• The same as the last outcome• The same as the last failure• The same as the last success• The same as an special old case with this
game• The same as an old case with another game• Perceived elements (context) guide the
reminding and analogy
MindRACES, First Review Meeting, Lund, 11/01/2006 6NBU
Examples of Anticipation based on analogy-making and context
• Predicting your partner’s or your rival’s next move
• What would I do in this situation (analogy with myself)
• What has this partner/rival done is analogous situation in the past (reminding of specific old case)
• What has another partner/rival done is analogous situation in the past (reminding of specific old case)
• Perceived elements (context) guide the reminding and analogy
MindRACES, First Review Meeting, Lund, 11/01/2006 7NBU
Analogy-Making
• Analogy-making is the transfer of a system of relations from one domain (base) to another (target). Similarity based on structure, not overall similarity.
• Analogy is a very basic human ability.
MindRACES, First Review Meeting, Lund, 11/01/2006 8NBU
Analogy-Making
in
water
tpot
in
oven
in
milkw
tpot
on
hplate
corr-to
corr-to
corr-to
corr-to
corr-to
MindRACES, First Review Meeting, Lund, 11/01/2006 9NBU
Rutherford’s Analogy
Sun Nucleus++
-
-
MindRACES, First Review Meeting, Lund, 11/01/2006 10NBU
Rutherford’s analogyRutherford’s analogy
sun
planet
yellow
mass
mass
temperature
greater
color
revolves-around
attractsTsun
Tplanet
Msun
Mplanet
causes
temperature
greater
mass
mass
revolves-around
attractsMnucleus
causes ?greater
nucleus
electron
Melectron
sun
planet
yellow
mass
mass
temperature
greater
color
revolves-around
attractsTsun
Tplanet
Msun
Mplanet
causes
temperature
greater
mass
mass
revolves-around
attractsMnucleus
causes ?greater
nucleus
electron
Melectron
The hydrogen atom is like our solar system.
The Sun has a greater mass than the Earth and attracts it, causing the Earth to revolve around the Sun. The nucleus also has a greater mass then the electron and attracts it. Therefore it is plausible that the electron also revolves around the nucleus.
MindRACES, First Review Meeting, Lund, 11/01/2006 11NBU
Main Implementation Tool - AMBR
• AMBR – a cognitive model of human analogy-making.
• The model is hybrid and integrates symbolic processing and connectionist spreading activation and constraint satisfaction at a micro level.
• The model is highly parallel and the behavior of the macro system emerges from the local interactions of micro agents.
MindRACES, First Review Meeting, Lund, 11/01/2006 12NBU
Challenges to the pre-existing version of AMBR
• AMBR was a theoretical tool – it was never applied in realistic domain before.
• AMBR was developed for complex problem-solving, not for anticipation.
• AMBR was a model of the mind outside of a body – no interactions with the environment – no perception, no manipulation.
• AMBR was coded in LISP with no possibilities for communications with other software.
MindRACES, First Review Meeting, Lund, 11/01/2006 13NBU
Scenario Implementation
• Selection of the scenarios to be used by NBU
• Developing simulation tools• First simulation experiments
MindRACES, First Review Meeting, Lund, 11/01/2006 14NBU
Scenarios studied by NBU
• Finding and Looking for an object (finding an object in a single room or in a maze of multiple rooms)
• Guards and thieves (collecting objects which are guarded by other agents)
MindRACES, First Review Meeting, Lund, 11/01/2006 15NBU
Rooms layout
Room A Room B
Room C
Room D Room E
Room F
Room G
N
MindRACES, First Review Meeting, Lund, 11/01/2006 16NBU
Looking for an Object (Scenario 1)
Room B
Room C
Room D Room E
Room F
Room G
N
MindRACES, First Review Meeting, Lund, 11/01/2006 17NBU
Guards and thieves (Scenario 3)
Room C
Room D Room E
Room F
Room G
N
Treasure hunters
Guards
MindRACES, First Review Meeting, Lund, 11/01/2006 18NBU
Developing Simulation Tools
• The AMBR model is being further developed and re-implemented in C#.
• The software for AIBO and Pioneer 3 is being mastered and tested.
• The simulation environment WEBOTS 5 is studied and simple simulation of the scenarios are being built.
• A middle tier is being implemented for communication between AMBR on one side and the robots and simulated environment on the other.
MindRACES, First Review Meeting, Lund, 11/01/2006 19NBU
Overall System Architecture
Overall System Architecture
Re
aso
ning
tie
rR
eas
oni
ng t
ier
Mid
dle
tier
Mid
dle
tier
Wo
rld ti
er
Wo
rld ti
er
Simulated or Real
Pre-processes data from other tiers
The core of the system
Real robots(AIBO, Pioneer, …)
Webots(Simulated AIBO,
Pioneer, …)
Processes data from and to the world layer.
Processes data from and to the reasoning layer.
DUAL/ AMBRThe analogical machine
Responsible for the inteligent/ anticipatory behaviour of the whole system
Web
ots
API
Mid
dle
tie
r API
Rob
ot s
pec
ific
API
Mid
dle
tie
r API
XM
L St
ruct
ure
d d
ata
WORLD
COMMUNI-CATION
REASONING
MindRACES, First Review Meeting, Lund, 11/01/2006 20NBU
World
AIBO ERS7 Webots simulation
MindRACES, First Review Meeting, Lund, 11/01/2006 21NBU
Communication
• World tier -> Reasoning tier Collect information about the world
using symbolic data from Webots Report it to the Reasoning layer
in suitable for AMBR form
• Reasoning tier -> World tier Get the motion plan from AMBR:
e.g “Go to the left cube” Send commands for movement to Webots
turning in place, walking forward
MindRACES, First Review Meeting, Lund, 11/01/2006 22NBU
Reasoning
• Reasoning by analogy with previous episode (using the AMBR cognitive model)
• Describing AMBR in UML• Implementation of the AMBR model in C#• Project infrastructure (version control, unit
testing, etc.)
MindRACES, First Review Meeting, Lund, 11/01/2006 23NBU
Anticipation by Analogy
?
MindRACES, First Review Meeting, Lund, 11/01/2006 24NBU
Past Episodes in Robot’s Memory
B1
cube
B2
Unique color
Same color
cylinders
B3
Left-of 3
Left-of 1
Left-of 2
prisms
B4
neighbor
pyramids
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
Target situation
MindRACES, First Review Meeting, Lund, 11/01/2006 25NBU
Results from the Simulation of Anticipation by Analogy
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B1
cube
MindRACES, First Review Meeting, Lund, 11/01/2006 26NBU
Results from the Simulation of Anticipation by Analogy
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B2
Unique color
Same color
cylinders
MindRACES, First Review Meeting, Lund, 11/01/2006 27NBU
Results from the Simulation of Anticipation by Analogy
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B3
Left-of 3
Left-of 1
Left-of 2
prisms
MindRACES, First Review Meeting, Lund, 11/01/2006 28NBU
Results from the Simulation of Anticipation by Analogy
B4
neighbor
pyramids
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
MindRACES, First Review Meeting, Lund, 11/01/2006 29NBU
Results from the Simulation of Anticipation by Analogy
Simulation2.1
BASES
B1 B2 B3 B4
Simulation2.1
TARGET
T1
Unique color
Same color
cube cylinders
Left-of 3
Left-of 1
Left-of 2
neighbor
prisms pyramids
balls
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B4
neighbor
pyramids
MindRACES, First Review Meeting, Lund, 11/01/2006 30NBU
Results from the Simulation of Anticipation by Analogy
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B2
Unique color
Same color
cylinders
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
B3
Left-of 3
Left-of 1
Left-of 2
prisms
B4
neighbor
pyramids
Same color
Left-of 1 Left-of 2
Unique color
Left-of 3
Neighbor 1 Neighbor 2
MindRACES, First Review Meeting, Lund, 11/01/2006 31NBU
Simulation Result - Video
Video Clip
MindRACES, First Review Meeting, Lund, 11/01/2006 32NBU
Challenges and Problems
• AMBR was developed as a model of complex analogies and therefore fitting and changes were required to produce anticipation: Superficial features such as colors are typically
ignored – colors are important in this domain; Episodes are complex and differ significantly
from each other – episodes are very similar in this domain.
MindRACES, First Review Meeting, Lund, 11/01/2006 33NBU
Challenges and Problems
• AMBR was developed as an isolated reasoning model – needs to be integrated into a complete cognitive system: Perceptual abilities need to be integrated that will
encode the target situation – perception of objects, properties and relations – this is solved in the simulation environment, needs to be solved with real robots; integration of symbolic and sub-symbolic approach;
Selective attention needs to be modeled to limit the representation of the target and to focus on certain aspects of the situation;
Motor control – planning and motor control mechanisms
MindRACES, First Review Meeting, Lund, 11/01/2006 34NBU
Challenges and Problems
• The simulation results need to be compared and possibly fitted to human data: Some of the simulation data perfectly match
human data; Some differ significantly.
MindRACES, First Review Meeting, Lund, 11/01/2006 35NBU
Comparing Simulation and Human Data: 100 Runs on each Target
А B C D
1 2 3 4 5 6 7 8
MindRACES, First Review Meeting, Lund, 11/01/2006 36NBU
Comparing Simulation and Human Data: 100 Runs on each Target
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4 5 6 7 8
A
Б
В
Г
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4 5 6 7 8
А
Б
В
Г
Simulation data Human data
MindRACES, First Review Meeting, Lund, 11/01/2006 37NBU
Integration with work of other partners
• Perception of objects, properties, relations – cooperation with IDSIA, LUCS, ISTC, OFAI
• Selective attention – integration of top-down and bottom-up mechanisms – cooperation with LUCS, IDSIA
• Emotions as regulators of the mechanisms of analogy-making, analogies as source of emotions – cooperation with ISTC, IST
MindRACES, First Review Meeting, Lund, 11/01/2006 38NBU
Anticipation by Analogy: Putting things together
Room A Room B
Room C
Room D Room E
Room F
Room G
N
Perception: target representation: IDSIA, LUCS
Motor control: OFAI, IDSIA, LUCS
Selective attention: LUCS, IDSIA, NBU
Emotions (IST, ISTC)
Analogical reasoning (NBU)
MindRACES, First Review Meeting, Lund, 11/01/2006 39NBU
Thank you for your attention!
?