From homeostasis to resource sharing: Biologically and economically compatible multi-objective multi-agent AI safety benchmarks

This work introduces safety challenges for an agent's ability to learn and act in desired ways in relation to biologically and economically relevant aspects. In total we implemented nine benchmarks, which are conceptually split into three developmental stages: "basic biologically inspired dynamics in objectives", "multi-objective agents", and "cooperation". The first two stages can be considered as proto-cooperative stages, since the behavioral dynamics tested in these benchmarks will be later potentially very relevant for supporting and enabling cooperative behavior in multi-agent scenarios. The benchmarks were implemented in a gridworld-based environment. The environments are relatively simple, just as much complexity is added as is necessary to illustrate the relevant safety and performance aspects.