Work

Kvantitatiivinen tutkimus vaikutuskykyä maksimoivan ja yleisen peliagentin vuorovaikutuksesta

Year: 2022
Type: dissertation
Abstract: Coupled empowerment maximisation (CEM) is an algorithm for implementing the behaviour of non-player characters (NPC) in video games. To this end, it employs empowerment, a computational intrinsic reward measuring an agent’s potential to influence the perceived environment. CEM drives the agent to maximise its own and maximise or minimise another agent’s empowerment while keeping the other agent within its influence. It has been hypothesised in previous qualitative research that CEM can drive adversarial and supportive NPC behaviour and generalises well amongst various discrete-time and discrete-space games. Integrating a CEM agent into a game requires no training and only minimal tweaks to the algorithm and game mechanics. The purpose of this thesis is to assess quantitatively whether different CEM parametrisations induce antagonistic or supportive behaviour. We recorded 39240 test runs in total, including the end score and game procession. In the test runs, a player agent controlled by an MCTS algorithm plays the test game with three different CEM agents in 120 different maps, with varying game rules. Firstly, we found that the CEM parametrisation, which we expected to yield antagonistic behaviour, did cause the player agent to score fewer points in our test game compared to an NPC that chose actions randomly. However, supportive behaviour did not emerge in our tests. Secondly, we show that the player agent and the NPC select actions with different frequencies depending on the CEM parametrisation. We conclude that changing the parametrisation can affect how and how well the player reaches their goals in the game and how they interact with the NPC. We also reflect on our findings based on observations of individual game runs.
Institution: Aalto University
Language: English
Open Access status: green
OpenAlex ID: W7117639690
Cites: 29
Cited by:
Related to: 9