Constructing Intelligent Agents via Neuroevolution By Jacob Schrum schrum2@cs.utexas.edu.

Constructing Intelligent Agents via Neuroevolution

By Jacob Schrum

schrum2@cs.utexas.edu

Motivation

• Intelligent agents are needed– Search-and-rescue robots– Mars exploration– Training simulations– Video games

• Insight into nature of intelligence– Sufficient conditions for emergence of:

• Cooperation• Communication• Multimodal behavior

Talk Outline

• Bio-inspired learning methods– Neural networks– Evolutionary computation

• My research– Learning multimodal behavior– Modular networks in Ms. Pac-Man– Human-like behavior in Unreal Tournament

• Future work• Conclusion

Artificial Neural Networks

• Brain = network of neurons

• ANN = abstraction of brain– Neurons organized into layers

Inputs Outputs

What Can Neural Networks Do?

• In theory, anything!– Universal Approximation

Theorem–

• Can’t program: too complicated• In practice, learning/training is hard

– Supervised: Backpropagation– Unsupervised: Self-Organizing Maps– Reinforcement Learning: Temporal-Difference

and Evolutionary Computation

MN ]1,0[]1,0[

Evolutionary Computation

• Computational abstraction of evolution– Descent with modification (mutation)– Sexual reproduction (crossover)– Survival of the fittest (natural selection)

• Evolution + Neural Nets = Neuroevolution– Population of neural networks– Mutation and crossover modify networks– Net used as control policy to evaluate fitness

Neuroevolution Example

Start WithParent Population

Neuroevolution Example

Start WithParent Population

Evaluate andAssign Fitness

100 90 75 61 56 50 31

Neuroevolution Example

Start WithParent Population

Evaluate andAssign Fitness

100 90 75 61 56 50 31

Clone, Crossoverand Mutate

To Get ChildPopulation

Neuroevolution Example

Start WithParent Population

Evaluate andAssign Fitness

100 90 75 61 56 50 31

Clone, Crossoverand Mutate

Children Are Nowthe New Parents

Repeat Process:Fitness Evaluations

As the process continues, each successive population improves performance

100 120 69 99 60 83 50

Neuroevolution Applications

F. Gomez and R. Miikkulainen, “2-D Pole Balancing With Recurrent Evolutionary Networks” ICANN 1998

Double Pole Balancing

Neuroevolution Applications

F. Gomez and R. Miikkulainen, “Active Guidance for a Finless Rocket Using Neuroevolution” GECCO 2003

Finless Rocket Control

Neuroevolution Applications

N. Kohl, K. Stanley, R. Miikkulainen, M. Samples, and R. Sherony, "Evolving a Real-World Vehicle Warning System" GECCO 2006

Vehicle Crash Warning System

Neuroevolution Applications

K. O. Stanley, B. D. Bryant, I. Karpov, R. Miikkulainen, "Real-Time Evolution of Neural Networks in the NERO Video Game" AAAI 2006

Training Video Game Agents

http://nerogame.org/

What is Missing?

• NERO agents are specialists– Sniping from a distance– Aggressively rushing in

• Humans can do all of this, and more

• Multimodal behavior– Different behaviors for

different situations

• Human-like behavior– Preferred by humans

What I do With Neuroevolution

• Discover complex agent behavior• Discover multimodal behavior

Contributions:• Use multi-objective evolution

– Different objectives for different modes

• Evolve modular networks– Networks with modules for

each mode

• Human-like behavior– Constrain evolution

Pareto-based Multiobjective Optimization

High health but did not deal much damage

Dealt lot of damage,but lost lots of health

Tradeoff between objectives

solution optimal one than More

)(

s.t. in points all contains

optimal Pareto is

:best points dominated-Non

)}(,,1{ 2.

and )}(,,1{ 1.

i.e. , dominates

RemainingHealth -

Dealt Damage -

:objectives with twogame Imagine

xyFyAx

FA

FA

uvni

uvni

uvuv

ii

ii

Non-dominated Sorting Genetic Algorithm II

• Population P with size N; Evaluate P• Use mutation (& crossover) to get P´ size N; Evaluate P´• Calculate non-dominated fronts of P P´ size 2N• New population size N from highest fronts of P P´

K. Deb, S. Agrawal, A. Pratap, T. Meyarivan, "A Fast Elitist Non-dominated sorting genetic algorithm for multi-objective optimization: NSGA-II" PPSN VI, 2000

Ms. Pac-Man

• Popular classic game• Predator-prey scenario

– Ghosts are predators– Until power pill is eaten

• Multimodal behavior needed– Running from threats– Chasing edible ghosts– More?

Modular Networks

• Different areas of brain specialize– Structural modularity → functional modularity

• Apply to evolved neural networks– Separate module → behavioral mode

• Preference neurons (grey)

arbitrate between modules• Use module with highest

preference output

( )( )

Module Mutation

• Let evolution decide how many modules

Networks start withone module

New modules addedby one of severalmodule mutations

Previous

Random

Duplicate

Intelligent Module Usage

• Evolution discovers a novel task division– Not programmed

• Dedicates one module to luring (cyan)

• Improves ghost eating when using other module

Comparison With Other Work

Authors Method Game AVG MAX

Alhejali and Lucas [1] GP FourMaze 16,014 44,560

Alhejali and Lucas [2] GP+Camps FourMaze 11,413 31,850

My Module Mutation Duplicate Results FourMaze 32,647 44,520

Brandstetter and Ahmadi [3] GP CIG 2011 19,198 33,420

Recio et al. [4] ACO CIG 2011 36,031 43,467

Alhejali and Lucas [5] GP+MCTS CIG 2011 32,641 69,010

My Module Mutation Duplicate Results CIG 2011 63,299 84,980

[1] A.M. Alhejali, S.M. Lucas: Evolving diverse Ms. Pac-Man playing agents using genetic programming. UKCI 2010.[2] A.M. Alhejali, S.M. Lucas: Using a training camp with Genetic Programming to evolve Ms Pac-Man agents. CIG 2011.[3] M.F. Brandstetter, S. Ahmadi: Reactive control of Ms. Pac Man using information retrieval based on Genetic Programming. CIG 2012.[4] G. Recio, E. Martín, C. Estébanez, Y. Sáez: AntBot: Ant Colonies for Video Games. TCIAIG 2012.[5] A.M. Alhejali, S.M. Lucas: Using genetic programming to evolve heuristics for a Monte Carlo Tree Search Ms Pac-Man agent. CIG 2013.

Types of Intelligence

• Evolved intelligent Ms. Pac-Man behavior– Surprising module usage– Evolution discovers the unexpected– Diverse collection of solutions

• Still not human-like– Human-like vs. optimal– Human intelligence

Modern Game: Unreal Tournament

• 3D world with simulated physics• Multiple human and software agents interacting• Agents attack, retreat, explore, etc.• Multimodal behavior required to succeed

Human-like Behavior: BotPrize

• International competition at CIG conference

• A Turing Test for video game bots– Judge as human over 50% of time to win– After 5 years, we won in 2012

• Evolved combat behavior– Constrained to

be human-like

Guessing Game

• Coleman: ????• Milford: ????• Moises: ????• Lawerence: ????• Clifford: ????• Kathe: ????• Tristan: ????• Jackie: ????

Judging Game

Player Identities

• Coleman: UT^2 (Our winning bot)• Milford: ICE-2010 (bot)• Moises: Discordia (bot)• Lawerence: Native UT2004 bot• Clifford: w00t (bot)• Kathe: Human• Tristan: Human• Jackie: Native UT2004 bot

Human Subject Study

• Six participants played the judging game

• Recorded extensive post-game interviews

• What criteria to humans claim to judge by?

Lessons Learned

• Don’t be too skilled– Evolved with accuracy restrictions– Disable elaborate dodging

• Humans are “tenacious”– Opponent-relative actions– Encourage “focusing” on opponent

• Don’t repeat mistakes– Database of human traces to get unstuck

Enem y

Bot

Item

Bot Architecture

Future Work

• Evolving teamwork– Ghosts must cooperate to eat Ms. Pac-Man– Unreal Tournament supports team play

• Domination, Capture the Flag, etc.

• Interactive evolution– Evolve in response to human interaction

• Adaptive opponents/assistants• Evolutionary art• Content generation

http://picbreeder.org/

Conclusion

• Evolution discovers unexpected behavior

• Modular networks learn multimodal behavior

• Human behavior not optimal– Evolution can be constrained to be

more human-like

• Many directions for future research

Questions?

contact

Jacob Schrum

schrum2@cs.utexas.edu