Presentation
in
Affinity Event: Indigenous in AI/ML
Mācistan: Reciprocity in Multi-Agent Reinforcement Learning as a Credit Assignment Problem
Dane Malenfant
Humans and animals have complex social and cultural systems that can span large distances and times. In North America, Plains Indigenous nations like the Métis Nation had a system of reciprocity involving effigies called manitohkan with food, tools and medicine for traveling. Leaving unneeded items at these can be perceived as cooperative actions; a reciprocal sharing of resources or caching of items. Mācistan is a novel MiniGrid environment built off the key-to-door task most common in temporal credit assignment reinforcement learning literature. However, abstracting to multi-agent dimension has led to interesting behaviour and the structural credit assignment problem.
Live content is unavailable. Log in and register to view live content