this is a condensed form of a presentation going over a lot of the ai workings in my game Sor. for a more standard presentation viewing experience, refer to the PDF form of the presentation.

ducks

decisions, planning, personality

note: project is wip

• everything in this project is highly wip

• any amount of the bird planning architecture may change

• sor is a cool game

  • more of a simulation than a game

• we don't use capitalization and highlight unnecessarily

decision modeling

utility ai

• how "useful" is an action?

• given a list of considerations

  • score each consideration by running various appraisals

  • transform inputs on curves to score levels

  • scores are in the range [0, 1]

  • optional thresholds attached to considerations

    • score is 0 if score < threshold

example: eat food

• consideration "eat"

  • appraisal "hunger"

    • a curve that sharply increases as energy percentage drops

  • appraisal "food availability"

    • discrete 0 or 1 for current availability of food

  • threshold 0.6

• plans the "eat" objective on success

reasoner architecture

• consideration

  • series of appraisals and an action

• appraisal

  • function returning a score

• reasoner

  • given a series of considerations, choose the "best"

    • the highest score is not necessarily the "best"

    • sometimes, to introduce unpredictability and variety, a fuzzy-decision strategy is used to choose between several similarly-scoring considerations

planning

goal-oriented action planning (GOAP)

• plan a sequence of actions to satisfy a goal

  • actions specify preconditions, postconditions and costs

  • goals are specified as a final postcondition

• graph search for optimal path

  • find the lowest-cost series of actions that reach the goal

  • an algorithm like A* can be used to tune search with a heuristic

  • A* heuristics can be biased to affect the "quality" of a plan

example: obtain food

• goal: satiety > TARGET_SATIETY

• actions:

  • eat bean

    • precondition: existence of bean

    • postcondition: satiety += BEAN_ENERGY

    • cost: distance/time to obtain bean

  • visit tree

    • precondition: existence of tree

    • postcondition: satiety += BEAN_ENERGY * TREE_BEANS

    • cost: distance/time to reach tree

how birds think

real-time bird thinking

• mind that runs a sense-think-act loop

  • a mind is associated with systems that can be run at different update rates

  • work offloaded to background threads

    • handle more computationally intensive planning

    • previous plans can be executed until new plans are available

• sense

  • reads information from the environment

• think

  • pick the most valuable objective and plan actions

• act

  • carry out real-time adjusted plans

phase 1: select objective

• use a reasoner to choose an objective

  • eat

    • get energy by consuming beans

  • explore

    • discover more of the map

  • defend

    • self-preservation by fight or flight

  • socialize

    • build opinion with other birds

phase 2: plan actions

• map the objective to action planning model

  • the phase 1 objective chooses the relevant actions

  • current state of the model based on sensed input

  • goal state is determined by the objective and bird state

• use GOAP planner to get the best sequence of actions

  • plan is a series of action nodes

  • each node contains an intermediate state of the model

interaction queue

• map abstract action plan to virtual controller input

• interaction plans represent queued actions in the world

  • target sources provide positional objectives

    • fixed locations such as trees

    • moving locations such as other birds or beans

  • approach types are used to keep distances

    • flocking via custom approaches on multiple birds

  • task plans represent physical actions

    • feed task involves distance tracking then feeding another bird

personality engine

original idea: lunch simulation

• simulate people eating lunch at restaurants

• each person has a personality

  • represented by the HEXACO model

    • 6-aspect model of human personality used in psychology

  • immutable values

• traits can be calculated from a personality

  • traits describe a person's expected response to a situation

  • values influenced by different aspects of personality

  • since personality is fixed, trait values are also fixed

• interaction effects based on participant traits

  • conversations between people can affect their opinions of each other

personality representation

• personality values are stored as a vector

• traits computed from personality and trait vectors

  • dot product for trait value in range [-1, 1]

• interactions use a function of the trait value

  • the trait value is used to map along a curve

  • trait values are distributed around 0

    • represents the relative value to the baseline

bird personalities

• personality aspects

  • anxiety

  • social awareness

  • emotionality

  • v1

• common traits

  • loyalty

  • aggression

  • wariness

  • inquisitiveness

  • sociability

example: sociability trait

• example bird personality (+0.8, 0.4)

  • high anxiety and medium social awareness

• using the sociability trait (-0.1, 1.0)

• sociability = 0.32

• factored into an appraisal for a socialize consideration

• used to adjust the opinion bonus for being fed

bird emotions

• transient values based on recent interactions

  • emotions fade toward 0 with time at a fixed decay rate

• can influence behavior to a lesser extent

  • traits provide primary influences on behavior

  • emotions can slightly shift the balance in either direction

  • recent actions can secondarily influence actions

• cause different responses to the same stimuli

• emotional sensitivity determined by traits

  • traits affect how much an action impacts emotions

conclusion

fun birds

• once i figure out how to make birds fun

  • i will let you know

• the game runs on windows, mac, and linux