Neuroevolution Arena: New Benchmark for Evaluating AI in Artificial Life Systems
A recent study published on arXiv (2608.10323) presents Neuroevolution Arena, a GPU-enhanced spatial ecology of neural-network cells with independent parameters, aimed at assessing trained controllers within competitive artificial-life frameworks. The research reveals that trained controllers can receive varying rankings when evaluated during training compared to ecological assessments, necessitating a nested evaluation protocol. This protocol integrates three specific update-and-inheritance regimes (EvoEvo, EvoRL, and RLRL) with two neural architectures, conducting 50,000 generations across three separate training runs for each condition. From the 18 runs, one elite-controller artifact is chosen for a frozen-evaluation design involving 198 computational tasks. Initial results suggest that RL-enabled regimes demonstrate superior training fitness compared to EvoEvo, yet ecological evaluations yield a different ranking. This study fills a significant gap in the evaluation of AI agents in dynamic, competitive settings, providing a more accurate measure than conventional training metrics.
Key facts
- Neuroevolution Arena is a GPU-accelerated spatial ecology of independently parameterized neural-network cells.
- The study uses a nested evaluation protocol with three update-and-inheritance regimes: EvoEvo, EvoRL, and RLRL.
- Two neural architectures are crossed with the regimes for 50,000 generations in three independent training runs per condition.
- 18 elite-controller artifacts are selected for an aligned-run frozen-evaluation design comprising 198 computational jobs.
- Pairwise effects average three seed-defined ecological contexts: two cooperation-permitting and one attack-permitting.
- RL-enabled regimes achieve higher recorded training fitness than EvoEvo.
- Pairwise outcomes show a different ranking under ecological evaluation compared to training fitness.
- The research is published on arXiv under the identifier 2608.10323.
Entities
Institutions
- arXiv