Transcript Lecture 13-2

CAP6938
Neuroevolution and
Artificial Embryogeny
Real-time NEAT
Dr. Kenneth Stanley
February 22, 2006
Generations May Not Always Be
Appropriate
• When a population is evaluated simultaneously
– Many are observable at the same time
– Therefore, entire population would change at
once
– A sudden change is incongruous, highly
noticeable
• When a human interacts with one
individual at a time
– Want things to improve constantly
Steady State GA: One Individual Is
Replaced at a Time
• Start by evaluating entire first generation
• Then continually pick one to remove, replace it
with child of the best
Start:
Evaluate All
f1
f2
f3
f4
f5
f6
f7
f8
2) Create
offpsring from
good parents
Repeat…
3) Replace
removed
individual
1) Remove poor individual
Steady State During Simultaneous
Evaluation: Similar but not Identical
• Several new issues when evolution is real-time
–
–
–
–
Evaluation is asynchronous
When to replace?
How to assign fitness?
How to display changes
Regular NEAT Introduces Additional
Challenges for Real Time
• Speciation equations based on generations
• No “remove worst” operation defined in
algorithm
• Dynamic compatibility thresholding assumes
generations
Speciation Equations Based on
Generations
How to Remove the Worst?
• No such operation in generational NEAT
• Worst often may often be a new species
– Removing it would destroy protection of
innovation
– Loss of regular NEAT dynamics
Dynamic Compatibility
Thresholding Assumes A Next
Generation
Real-time NEAT Addresses Both the
Steady State and Simultaneity Issues
• Real-time speciation
• Simultaneous and asynchronous
evaluation
• Steady state replacement
• Fast enough to change while a game is
played
• Equivalent dynamics to regular NEAT
Main Loop (Non-Generational)
Choosing the Parent Species
Finally: How Many Ticks Between
Replacements?
•
Intuitions:
– The more often replacement occurs, the fewer are eligible
– The larger the population, the more are eligible
– The high the age of maturity, the fewer are eligible
rtNEAT Is Implemented In NERO
• Download at http://nerogame.org
• rtNEAT source soon available (TBA)
• Simulated demos have public appeal
– Over 50,000 downloads
– Appeared on Slashdot
– Best Paper Award in Computational Intelligence and
Games
– Independent Games Festival Best Student Game
Award
– rtNEAT licensed
– Worldwide media coverage
NERO: NeuroEvolving Robotic
Operatives
• NPCs improve in real time as game is played
• Player can train AI for goal and style of play
• Each AI Unit Has Unique NN
NERO Battle Mode
• After training, evolved behaviors are saved
• Player assembles team of trained agents
• Team is tested in battle against opponent’s
team
NERO Training: The Factory
• Reduces noise during evaluation
– All evaluations start out similarly
• Robot bodies produced by “factory”
• Each body sent back to factory to respawn
• Bodies retain their NN unless chosen for
replacement
• NN’s have different ages
– Fitness is diminishing average of spawn trials:
NERO Inputs and Outputs
Enemy/Friend Radars
Enemy On-Target Sensor
Object Rangefinder Sensors
Enemy Line-of-Fire Sensors
Further Applications?
•
•
•
•
•
New kinds of games
New kinds of AI in games
New kinds of real-time simulations
Training applications
Interactive steady-state evolution
Next Topic:
Improving the neural model
• Adaptive neural networks
• Change over a lifetime
• Leaky integrator neurons and CTRNN
Evolutionary Robots with On-line Self-Organization and Behavioral Fitness by Dario
Floreano and Joseba Urzelai (2000)
Evolving Adaptive Neural Networks with and Without Adaptive Synapses by Kenneth O.
Stanley, Bobby D. Bryant, and Risto Miikkulainen (2003)
New Homework due 3/8/06 (see next slide)
Homework Due 3/8/06
Genetic operators all working:
•Mating two genomes: mate_multipoint, mate_multipoint_avg, others
•Compatibility measuring: return distance of two genomes from each other
based on coefficients in compatibility equation and historical markings
•Structural mutations: mutate_add_link, mutate_add_node, others
•Weight/parameter mutations: mutate_link_weights, mutating other parameters
•Special mutations: mutate_link_enable_toggle (toggle enable flag), etc.
•Special restrictions: control probability of certain types of mutations such as
adding a recurrent connection vs. a feedforward connection
Turn in summary, code, and examples demonstrating that all
functions work. Must include checks that phenotypes from
genotypes that are new or altered are created properly and work.
Project Milestones (25% of grade)
•
•
•
•
•
•
2/6: Initial proposal and project description
2/15: Domain and phenotype code and examples
2/27: Genes and Genotype to Phenotype mapping
3/8: Genetic operators all working
3/27: Population level and main loop working
4/10: Final project and presentation due (75% of grade)