SIMULATING A FREEZING SOCIETY BUILDING THE AI OF FROSTPUNK HELLO - - PowerPoint PPT Presentation

SIMULATING

A FREEZING SOCIETY BUILDING THE AI OF FROSTPUNK

HELLO

I’M MACIEJ CZERWONKA GAMEPLAY PROGRAMMER | SIMULATION TEAM GAME DEVELOPER FOR 15 YEARS @Calveit

WE ARE HIRING

FROSTPUNK AI BUILDING BLOCKS PERFORMANCE

FROSTPUNK

SIMULATION

AI BUILDING BLOCKS

AGENTS

Agent

• Core simulation element
• Represents a citizen
• Interacts with environment

State

• Agent memory and needs
• Private set of data
• Accessable only for decision

tree

Perception

• Eyes and ears of the agent
• Translates environment to

format usable by decision tree

Decision tree

• Agent brain
• Has access to state and

perception

• Behaviour tree
• Affects the state

ACTIVITIES

Activity

• Verb of a simulation
• Not atomic
• Not an action

Activity tree

• Action factory (produce actions

for agents)

• Behaviour tree
• Multiple agents allowed

Target selector

• Matcher (assigns agents to
• bjects)

Priority

• Number
• Decides the order
• Easy to change

! Hard to maintain

Activity list

• One dimensional list
• Decides the order
• Transparency
• Visibility
• Cheap maintenance

Preferences

• Set of values
• Agent decision
• Fallback activities always in the

collection

Action

• Controlling agent
• Can influence state
• Can influence visuals

Plan

• Collection of actions
• FIFO
• Fed by activity tree
• Consumed by agent when

current action is finished

POINT OF INTEREST

Point of interest

• POI
• Agents perform activities here
• Attached to building or other

place in the game

• Place can have more than 1 POI

Activity

• Activity type available at POI
• Each POI declares an activity

Conditions

• Collection of expressions
• Evaluates to boolean value,

which decides if active

Space

• Range (eg. 1-10)
• How many agents can perform

activity here

Target

• Agent always has a target
• Binds agent with POI
• Idle agents have target
• Dead agents have target

Is this model good enough?

• Doctors leave their patients to build a road
• Children fight in Fight Club
• Hunters return from 3-day expedition to get a snack
• Guards join the protest, they were supposed to disperse
• Prison workers leave their work to bury the dead, letting convicts escape

No, it is not. We need to distinguish agents!

AGENTS

AGENTS

AGENTS

AGENTS

Features

• Collection of Label objects
• Label is just a string, can be

anything

• Examples: OnExpedition, Female,

Engineer, Disabled, etc.

POI spaces

• Collection of Space objects
• Count below minimum is

forbidden

• Count over maximum is

forbidden

• Only agents with labels in

Required collection are accepted

• Agents with labels present on

Prohibited collection are denied

We have model. What now? Let’s assign agents to POIs.

How to assign agents?

Assignment problem

The assignment problem deals with assigning machines to tasks, workers to jobs, soccer players to positions, and so on. The goal is to determine the optimum assignment that, for example, minimizes the total cost or maximizes the team effectiveness.

Reference: http://www.hungarianalgorithm.com

Hungarian algorithm

• Solves assignment problem,
• Developed by Harold Kuhn in 1955,
• Based on work of Hungarian mathematicians

Dénes Kőnig and Jenő Egerváry (thus the name Hungarian),

• Fairly easy to understand,
• Our implementation is based on Topcoder

article.

Reference: http://www.convexoptimization.com

Assignment problem - example

Reference: http://www.hungarianalgorithm.com

Assignment problem - example

Reference: http://www.hungarianalgorithm.com

Formulating assignment problem

COST FUNCTION? DISTANCE

One activity at a time

One activity at a time

Occupied POIs

Rebalancing

Invalid assignment POI

Minima satisfaction

Minima satisfaction algorithm

Hungarian algorithm

Eliminate POIs with broken minima

Sort spaces by saturation level

Saturate spaces

Results

Targeting overview

• Determine a target for each agent
• Centralized process, not an agent decision
• One activity at a time
• All agents at a time
• All POIs at a time
• Rebalancing
• Satisfy POI minima
• Kicking an agent is also targeting
• Agent death is also targeting

We have model. We have targeting. Are we good?

Let’s see

• People change houses every day
• Families do not stay together
• People change work every day
• Patients who are almost cured get kicked out of medical posts

Why does it happen? Distance is not universal cost descriptor!

Target selectors

• Function which matches agents with POIs
• Activity may want to override default target selector
• Target selector should have its own cost function
• Target selector should decide if it wants rebalancing
• Transparent for targeting

Target selector examples

• NEAREST
• Cost descriptor: distance
• No rebalancing
• Use case: construction
• PERSISTENT ASSIGNMENT
• Cost descriptor: assignment
• No rebalancing
• Use case: work
• MEDICAL TREATMENT
• Cost descriptor: distance, but heal the almost cured in the first place, prefer patients who are

gravely ill

• Rebalancing
• Use case: healing

Does it work?

• Mass protests are attended by 3 people
• Convict decides not to show up on his execution
• Idle citizens just „stand” there
• Production output is unpredictable

Living city still miss reliability Smoke and mirrors!

IDLE BEHAVIOURS

DIRECTED SCENES

INDEPENDENT ECONOMY

INDEPENDENT ECONOMY

INDEPENDENT ECONOMY

PERFORMANCE

Event Driven Behaviour Trees

Budgeting and distributed update

Parallel targeting

Limit number of processed entities

• Update only dirty agents

○ Dirty - kicked out of its POI or changed its preferences since last update

• Update only dirty points of interest

○ Dirty - constructed, upgraded, targeted/left by an agent, state changed

• Not possible with rebalancing

Lessons learned

From the very beginning of development, we focused too much on the simulation, failing to achieve living city reliability and economy stability. What we really wanted was not simulation, but experience. We wanted the player to believe that what he/she sees is a city populated with people who behave in reliable manner. We also wanted to have stable economy, which is a core system for a city-builder survival game. Doing this by parametrizing simulation is hard. Having said that, we should not have put most of our effort into the simulation. We should have focused more on player perception of the living city.

THANK YOU