Skip to content

Latest commit

 

History

History
11 lines (10 loc) · 1.63 KB

generative models.md

File metadata and controls

11 lines (10 loc) · 1.63 KB

Modeling players

  1. Generative Agents for Player Decision Modeling in Games
    • This paper describes an approach to modeling, grouping and interpreting players based on their inferred utility function modeled through reinforcement learning in the form of a generative agent.
    • Uses trained Q-learning agents with manually configured reward parameters as a priori defined personas.
    • Player modelling: description, prediction, interpretation, and in some cases reproduction of player behavior
    • Dungeons were generated via constrained genetic algorithms
    • Assumptions: The first assumption is that players exhibit a particular de-cision making tendency or style when playing a particularlevel or game, and that this tendency can be captured andexpressed by approximating autility functionthat shapestheir decisions in-game
    • Comments: Showing summary scrren of previous level, limited the notion of "best" to some combination of the features shown in the summary.
    • For each player's play-trace, we replay the whole game and at each point in time,we input the state description to all of our artificial agents,and compare the player’s decision to the decision of the dif-ferent agents. the met-ric was calculated as the number of agent-persona/humanplayer agreementsNafor each decision made in the humandecision trace, normalized with respect to the number of de-cisions in the player’s decision traceN, i.e.Na/N
    • Future work: Personas may change over time. A person who is collecting all treasures at the start, may start taking shortest path to exit if they get bored.