1 / 26

Evolution Neural Network Agents in the NERO Video Game

Evolution Neural Network Agents in the NERO Video Game. Advisor : Dr. Hsu Presenter : Chien-Shing Chen Author: Kenneth O. Stanley Bobby D.Bryant Risto Miikkulainen. CON. IEEE Symposium on Computational Intelligence and Games, April 2005. Outline. Motivation

Download Presentation

Evolution Neural Network Agents in the NERO Video Game

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Evolution Neural Network Agents in the NERO Video Game Advisor :Dr. Hsu Presenter: Chien-Shing Chen Author: Kenneth O. Stanley Bobby D.Bryant Risto Miikkulainen CON. IEEE Symposium on Computational Intelligence and Games, April 2005

  2. Outline • Motivation • Objective • Introduction • Experimental Results • Conclusions • Personal Opinion

  3. Motivation • game market • Game characters can learn through interacting with the player, keeping it interesting. • autonomous computer-controlled agent in the game • intelligent agents • Adapt and change in real-time • the behavior of agents in current games is often repetitive and predictable

  4. Objective • In NERO, the player takes the role of a trainer, teaching skills to a set of intelligent agents controlled by rtNEAT. • agent behavior improves visibly during gameplay

  5. Introduction • Real-time NEAT(rtNEAT) is able to complexifyneural networks as the game is played, making it possible for agents to evolve increasingly sophisticated behaviors in real time. • Agent behavior improves visibly during gameplay. • NEAT, rtNERO

  6. Background • NERO need to learn online as the game is played, predetermined training targets are usually not available. • Traditional techniques have posed significant challenges.

  7. Background 1. Large state/ action space • High-dimensional, real-time game 2. Diverse behaviors • same behavior in a homogeneous population 3. Consistent individual behaviors • players don’t want to see an individual agent periodically… 4. Fast adaptation • don’t want to wait hours for agents to adapt 5. Memory of past states • agents remember past events

  8. rtNEAT • if agents could be evolved in a smooth cycle of replacement, the play could interact with evolution during the game and the many benefits

  9. rtNEAT • A technique for evolving neural networks for complex reinforcement learning tasks using a genetic algorithm(GA) • Online, real-time, interact • NEAT is based on three key ideas: • Historical markings • Speciation(物種形成) • Starting from minimal structure

  10. rtNEAT

  11. rtNEAT 1. Historical markings • Evolving network structure requires a flexible genetic encoding • Each genome includes a list of connection genes, each of which refers to two node genes being connected. • Each connection gene specifies the in-node, the out-node, the connection weight, and an innovation number, which allows finding corresponding genes during corssover. • Mutation can change both connection weights (perturb) and network structures (add a new connection or a new node to the network). • Through mutation, genomes of varying sizes are created. • Each unique gene in the population is assigned a unique innovation number, and the numbers are inherited during crossover.

  12. rtNEAT 2. Speciation(物種形成) • 物競天擇適者生存 • 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

  13. rtNEAT 3. Starting from minimal structure • New structure is introduced incrementally as structural mutations occur, and only those structures survive that are found to be useful through fitness evaluations. • This way, NEAT searches through a minimal number of weight dimensions and finds the appropriate complexity level for the problem.

  14. Running NEAT in Real Time • Remove the agent with the worst adjusted fitness from the population assuming one has been alive sufficiently long so that is has been properly evaluated. • too old, too young

  15. Running NEAT in Real Time 2. Re-estimating • Was an agent old enough to be removed, its species now has one less member and therefore its average fitness has likely changed • used in choosing the parent species in the next step

  16. Running NEAT in Real Time 3. Choosing the parent species

  17. Running NEAT in Real Time 4. Dynamic Compatibility Thresholding 5. Replacing the old agent with the new one 6. Determining Ticks Between replacements

  18. Running NEAT in Real Time

  19. NeuroEvolving Robotic Operatives(NERO) • The robots begin the game with no skills and only the ability to learn.

  20. Training Mode

  21. Training Mode • Training begins by deploying 50 robots on the field. • Each robot is controlled by a neural network with random connection weights and no hidden nodes, as is the usual starting configuration for NEAT.

  22. Playing NERO

  23. Playing NERO

  24. Playing NERO

  25. Playing NERO

  26. Conclusions • Drawback • evaluation • Application • game domain • Future Work • implement

More Related