Evolving Neural Network Agents in the NERO Video Game

Author： Kenneth O. Stanley, Bobby D. Bryant, and Risto Miikkulainen

Presented by Yi Cheng Lin

Outline

Introduction The behavior of agents Challenges to traditional Reinforcement learn

ing (RL) techniques Real-time NeuroEvolution of augmenting Top

ologies (rtNEAT) NeuroEvolving Robotic Operatives (NERO) Playing NERO Conclusion

Introduction

The world video game market in 2002 was between $15 billion and $20 billion

This paper introduces the real-time NeuroEvolution of Augmenting Topologies (rt-NEAT)

It’s purpose is let Non-player-character (NPC) interact with palyers in game playing

The behavior of agents

The behavior of agents in current games is often repetitive and predictable

Machine learning could potentially keep video games interesting by allowing agents to change and adapt

a major problem with learning in video games is that if behavior is allowed to change, the game content becomes unpredictable

Challenges to traditional Reinfor -cement learning (RL) techniques

Large state/action space Diverse behaviors Consistent individual behaviors Fast adaptation Memory of past states

Real-time NeuroEvolution of augmenting Topologies (rtNEAT)

The rtNEAT method is based on NEAT, a technique for evolving neural networks for complex reinforcement learning task using a genetic algorithm

NEAT is based on three key idea

NEAT

First, tracking genes with historical markings to allow easy crossover between different topologies

each unique gene in the population is assigned a unique innovation number, and the number are inherited during crossover

protecting innovation via speciation

NEAT

Second, the reproduction mechanism for NEAT is explicit fitness sharing, where organisms in the same species must share the fitness of their niche, preventing any one species from taking over the population

Third, NAET begins with a uniform population of simple networks with no hidden nodes

Running NEAT in Real Time

rtNEAT

After every n ticks of the game clock, rtNEAT performs the following operation:

Step 1: Remove the agent with the worst adjusted fitness from the population assuming one has been alive sufficiently long so that it has been properly evaluated

It is also important not to remove agents that are too young

rtNEAT

Step 2: Re-estimate F for all species (F : average fitness)

Step 3:Choose a parent species to create the new offspring ,where is the average fitness of species k, is the sum of all the average species fitness

rtNEAT

Step 4: Adjust compatibility threshold Ct dynamically and reassign all agents to species– the advantage of this kind of dynamic compatibilit

y thresholding is that it keeps the number of species relatively stable

Step 5: Replacing the old agent with the new one

Determining Ticks Between Replacements

The appropriate frequency can be determined through a principled approach

Parameter:– n : the ticks between replacements– I : the fraction of the population that is too young

and therefore cannot be replaced– m : is the minimum time alive– |P| is the population size

Determining Ticks Between Replacements

It is best to let the user choose I because in general it is most critical to performance

rtNEAT can determine the correct number of ticks between replacements n to maintain a desired eligibility level.

In NERO, 50% of the population remains eligible using this technique

NeuroEvolving Robotic Operatives (NERO)

Training Mode– The player sets up training exercises by placing o

bjects on the field and specifying goals through several sliders

Battle Mode

Avoiding turret fire

Navigating a maze

Conclusion

A real-time version of NEAT (rtNEAT) was developed to allow users to interact with evolving agents

Using this method, it was possible to build an entirely new kind of video game, NERO, where the characters adapt in real time in response to the player’s actions