Game Balancing with Ecosystem Mechanism
Download as PPTX, PDF0 likes111 views
The document discusses game balancing techniques utilizing ecosystem mechanisms, focusing on both offline and online strategies such as data collection, adaptive behavior of agents, and procedural content generation. It highlights challenges in balancing evolving behaviors and content, the importance of diversity in game experiences, and the heavy workload involved in analyzing player logs. Future work aims to improve adaptivity and handle more complex learning scenarios in game environments.
![Architecture
Particle containing Weights
[Number and types of
Game Objects]
16
(Behaviour)
(Content)
Single Swarm
Swarm
Particle containing
behaviour parameters
of a game object in
the swarm
Weight
* EIM – Ecosystem Implementing Model (a Swarm)](https://image.slidesharecdn.com/sapience-gamebalancinganand-170902170559/85/Game-Balancing-with-Ecosystem-Mechanism-16-320.jpg)
Game Balancing with Ecosystem Mechanism
- 1. Game Balancing with Ecosystem Mechanism 1 Wen Xia, Bhojan Anand (presenter)
- 2. Background 2
- 3. Balance Control 3 Too easy Too hard Time/Skill Difficulty
- 4. State of Art Approaches for Game Balancing Offline Strategies • Pre-defined behaviors for agents • Manual adjustment and update game versions • Data collection and offline learning • Quantitative methods • Etc. 4
- 5. State of Art Approaches for Game Balancing Cont. Online/Real-time Strategies - DDA (most common) • Adaptive Behavior of Agents (power, size, attack frequency …) • PCG (Procedural content generation, eg more weapons, new weapons, buildings ) 5
- 6. Online Strategies Adaptive Behavior • Dynamic scripting • Real-time genetic control • PSO-ANN model • Etc. 6
- 7. Online Strategies Cont. PCG • Experience-driven: personalized content • Search-based: generate and test • Learning-based: modeling game with content features • Etc. 7
- 8. Problems Evolution on behavior and content simultaneously is seldom considered • Evolving either one will influence the performance of the other one in total outcome. • Integration of different learning strategies will cause developers extra work. Diversity on results is ignored • Diversity on contents and behaviors creates curiosity and fun. • Players usually do not notice the little difference among “best solution” and “good solutions”. Analyzing player-log causes heavy workload. • Mapping from player-log to features, and from features to game content causes too much manual work. 8
- 11. Ecosystem & Game Environment 11 Similarities • Diversity of agents with simple behavior rules. • Different types of agents may have strength gaps. • Need to be balanced. Differences • Ecosystem agents have complicated relationships among each other. • Game agents mostly interact with player’s character.
- 12. Ecosystem Macro-mechanism Balance point • Resource limitation Cycle chain • Different species play different roles Diversification • Same population, not always same behavior 12
- 13. Ecosystem Macro-mechanism Cont. Balance point • Resource limitation Cycle chain • Different species play different roles Diversification • Same population, not always same behavior 13 Total balance goal Individual balance goals Action weight diversity
- 14. Ecosystem Micro-mechanism Simple behavior for one individual • What to eat, how to move, etc. Local perception • Self knowledge, neighbor communication Evolution • Next generations will be evolved 14
- 15. Ecosystem Micro-mechanism Cont. Simple behavior for one individual • What to eat, how to move, etc. Local perception • Self knowledge, neighbor communication Evolution • Next generations will be evolved 15 Swarm intelligence
- 16. Architecture Particle containing Weights [Number and types of Game Objects] 16 (Behaviour) (Content) Single Swarm Swarm Particle containing behaviour parameters of a game object in the swarm Weight * EIM – Ecosystem Implementing Model (a Swarm)
- 17. Diversity Rationale All swarms and agents (objects) have random initial positions (including newly born agents). Different swarms may result in different local optima. Stable agent will randomly mutate (low possibility). 17 Evolution Space
- 18. Diversity Rationale Cont. Example of one set of weights of “bot” agent • 3 swarms, 3 action weights • Agents within same swarm have similar but not same weights • Some swarms contain randomly mutated agents (6th of top, 1st of bottom) 18
- 19. Learning Flow n, k are manually settled. Stochastic environment → n++ Large action base → k++ 19 * Goals in Level1 and Level 2 should be aligned with the ‘Game Balancing’ factor
- 20. Goals Player’s Power Variables • Objective variables: HP, armor, weapon strength, character level, etc. • Subjective variables: proficiency, habit, emotional condition, etc. • Networking variables: team cooperation/competition. System Requirements • Set goals within evolution range manually. • Balancing Goal (eg, based on objective variables) • Full understanding of game content/agents/actions. 20 Future works Current goal
- 21. Test Implementation & Evaluation 21
- 22. Test Scene 22 Numerical Settings Demo Game settings HP remain Agent number of each types Learning record Individual performanceGame settings
- 23. Evaluation Objectives 1) Adaption We set different maximum HP and armor of the player, in order to prove that our approach is able to adapt when the properties of the player change. 2) Diversification We need to prove that the results of behaviours and content types will not always be similar if we repeat same test cases again. 23
- 24. Adaptivity Results (armor) Final error: ± 5% Passed Results: 8 Failed Results: 1 • 1B • Caused by early convergence of PSO 24 Repeated, A-C have same conditions
- 25. Adaptivity Results (HP) Final error: ± 5% Passed Results: 9 Observation: higher HP, more stable learning curve • Same amount of object number change, less influence on total error. 25
- 26. Diversity Results (content & behavior) 26
- 27. Complete Results of 18 Tests https://docs.google.com/spreadsheets/d/1U05mqhuwOBv71ieFPqBjkTK2JHE04w 1fqc6TcJNSwx4/edit#gid=229731456 27
- 28. Disadvantages & Future Work Not so fast • Everytime when updated, need to test again by at least one iteration/generation to collect feedback from game environment. Large error in first several iteration/generations • Player might feel weird. Cannot handle aggregation • Advanced learning algorithm might support. 28
- 29. Q&A 29
- 30. MORE SLIDES & DEMO – for Q&A Support! 30
- 31. Test Settings for Learning Component 31
- 32. Test Settings for Game Objects 32
- 33. Test Cases 33
- 34. Balance Goals 34 Bots 30 damage / sec Ghost 60 damage / sec Succubus 110 damage / sec Player’s HP at the end of one generation 0 Back to Test Scene
- 35. Demo – Game Balancing
- 36. v: velocity (disposition), x: position (state), p: best position, t: current time step, i: individual, g: neighbor, ω: inertia weight, c: cognitive weight, r: random number ∈ (0, 1) Particle Swarm Optimization (PSO) 36
- 37. PSO Cont. 37
- 38. Variances of PSO Multi-swarm Optimization Hierarchical Swarm Optimization Waves of Swarm Particles ... 38 Back to System Design
Editor's Notes
- #5: Pre-defined behaviors for agents - menu system ‘novice’, ‘player’, ‘expert’ - multiple levels (increasing difficulty….) Manual adjustment and update game versions - address balancing issues based on comments in new versions Data collection and offline learning - learn about player and tune the next version of game Quantitative methods - check the score and adapt difficulty according to score Etc.
- #17: The system design is constructed into two levels. The level 1 is for individual EIMs, which aims at evolving individual behaviors. The level 2 is for EIM manager, which aims at evolving the numbers and types of game objects. The EIM stands for ecosystem implementing model. One EIM contains one type of game objects, and it works as a swarm where the parameters that control behaviors of game objects are its particles. In level 1, same type of game objects might be splitted into different EIMs, such that even though they have same optimization goals, the final outcome could be totally different. In level 2, there is only one swarm, of which every particle inside contains one set of the numbers that correspond to the numbers of different types of game objects.
- #20: One learning cycle of individual EIMs is one iteration, and one learning cycle of EIM manager is one generation. One generation contains n iterations. And k is used to trigger the learning process of EIM manager. Because when the individual behavior is too noisy, there is meaning to evolve the number or type of game objects. If the game environment is very stochastic, n needs to be high, because it needs a longer period to reduce the possibility of extreme data. And if the action base of a game object is very large, k needs to be high, because the larger the searching space is, with the same acceptable error range, the longer time it needs for the particles to gather around the optimized value. Besides, currently, both n and k are manually settled. {Balancing Goal: - The balancing goal of enemies is to attack the player such that when one game round (one generation) ends, the HP of player will become 0. NOTE: one generation has a fixed time duration (in test cases: 10 secs x 10 iterations = 100 secs) - Bots (30 damage / sec) - Ghost (60 damage / sec) Succubus (110 damage / sec) }
- #21: {eg. Balancing Goal: - The balancing goal of enemies is to attack the player such that when one game round (one generation) ends, the HP of player will become 0. NOTE: one generation has a fixed time duration (in test cases: 10 secs x 10 iterations = 100 secs) - Bots (30 damage / sec) - Ghost (60 damage / sec) Succubus (110 damage / sec) } Evaluation objectives: 1) Adaption We set different maximum HP and armor of the player, in order to prove that our approach is able to adapt when the properties of the player change. 2) Diversification We need to prove that the results of behaviours and content types will not always be similar if we repeat same test cases again.