Authors:
Matthew Stephenson, Alex Newcombe, Eric Piette, Dennis J. N. J. Soemers

Venue:
IEEE Access, 2026

Topics:
general game playing, multi-armed bandits, best arm identification, agent evaluation, portfolio agents

Links: IEEE · PDF

This paper introduces an efficient and general procedure for identifying the best, or near-best, performing agent for each game in a general game playing benchmark.

The approach formulates the problem as a multi-bandit best arm identification task, where each game is treated as a bandit and each available agent as one of its arms. The authors propose a new algorithm called RCP (Regret Change Potential), which uses confidence intervals to prioritise trials that have the highest potential to reduce overall simple regret.

Experiments on both the GVGAI framework and the Ludii general game system show that RCP significantly outperforms previous state-of-the-art best arm identification algorithms, improving both average simple regret and average probability of error. The paper also reports strong gains in the early and middle stages of evaluation, where efficient trial allocation matters most.

Evaluating agents in general game playing is expensive because many games are stochastic, many agents require repeated trials to obtain reliable estimates, and modern frameworks contain hundreds or even thousands of games.

This paper addresses that challenge by shifting the focus from exhaustive evaluation to efficient identification of a strong agent for each individual game. This is especially useful for tasks such as building portfolio agents, selecting specialised controllers, and identifying gaps in an existing agent suite.

More broadly, the work contributes to the methodology of general game AI by improving how researchers compare agents across large and diverse benchmarks, and it is directly relevant to both GVGAI and Ludii-based research. The empirical results shown in the figures on pages 11 to 13 also make clear that the proposed RCP method consistently reduces regret faster than alternative approaches across both domains.

Stephenson, M., Newcombe, A., Piette, E., Soemers, D. J. N. J. (2026). Best Agent Identification for General Game Playing. IEEE Access.

@article{stephenson2026best_agent_identification,
  author  = {Stephenson, Matthew and Newcombe, Alex and Piette, Eric and Soemers, Dennis J. N. J.},
  title   = {Best Agent Identification for General Game Playing},
  journal = {IEEE Access},
  year    = {2026},
  note    = {Accepted},
  doi     = {10.1109/ACCESS.2026.3685576}
}