Transcript Lecture 2

CAP6938
Neuroevolution and
Artificial Embryogeny
Basic Concepts
Dr. Kenneth Stanley
January 11, 2006
We Care About Evolving Complexity
So Why Neural Networks?
•
•
•
•
Historical origin of ideas in evolving complexity
Representative of a broad class of structures
Illustrative of general challenges
Clear beneficiary of high complexity
How Do NNs Work?
Output
Output
Input
Input
How do NNs Work?
Example
Outputs (effectors/controls)
Forward Left Right
Front Left Right Back
Inputs (Sensors)
What Exactly Happens Inside the
Network?
• Network Activation
out1
out2
H1
H2
w11
w22
w12
X1
w21
X2
Neuron j activation:
 n

H j     xi wij 
 i 1

Recurrent Connections
• Recurrent connections are
backward connections in
Recurrent connection
the network
Wout-H
• They allow feedback
• Recurrence is a type of
memory
out
wH-out
H
w11
X1
w21
X2
Activating Networks of Arbitrary
Topology
• Standard method makes no distinction between
feedforward and recurrent connections:
H j , H j (t )
 n

    xi (t 1) wij 
 i 1

• The network is then usually activated once per
time tick
out
Wout-H
wH-out
• The number of activations per tick can be
H
thought of as the speed of thought
w11 w21
• Thinking fast is expensive
X
X
1
2
Arbitrary Topology Activation
Controversy
• The standard method is not necessarily the best
• It allows “delay-line” memory and a very simple
activation algorithm with no special case for
recurrence
• However, “all-at-once” activation utilizes the
entire net in each tick with no extra cost
• This issue is unsettled
The Big Questions
• What is the topology that works?
• What are the weights that work?
?
?
?
? ?
?
?
?
?
? ?
? ?
Problem Dimensionality
• Each connection (weight) in the network is a
dimension in a search space
21-dimensional space 3-dimensional space
• The space you’re in matters: Optimization is not
the only issue!
• Topology defines the space
High Dimensional Space
is Hard to Search
• 3 dimensional – easy
• 100 dimensional – need a good
optimization method
• 10,000 dimensional – very hard
• 1,000,000 dimensional – very very hard
• 100,000,000,000,000 dim. – forget it
Bad News
• Most interesting solutions are high-D:
– Robotic Maid
– World Champion Go Player
– Autonomous Automobile
– Human-level AI
– Great Composer
• We need to get into high-D space
A Solution (preview)
• Complexification: Instead of searching
directly in the space of the solution, start in
a smaller, related space, and build up to
the solution
• Complexification is inherent in vast
examples of social and biological progress
So how do computers optimize
those weights anyway?
• Depends on the type of problem
– Supervised: Learn from input/output examples
– Reinforcement Learning: Sparse feedback
– Self-Organization: No teacher
• In general, the more feedback you get, the
easier the learning problem
• Humans learn language without
supervision
Significant Weight Optimization
Techniques
• Backpropagation: Change weights based
on their contibution to error
• Hebbian learning: Changes weights based
on firing correlations between connected
neurons
Homework:
-Fausett pp. 39-80 (in Chapter 2)
-and Fausett pp. 289-316 (in Chapter 6)
-Online intro chaper on RL
-Optional RL survery