Package ai.badmonkey.agentspaces.capabilities.learn

package ai.badmonkey.agentspaces.capabilities.learn
  • Class
    Description
    GossipLearner<M>
    The typed aspace:cap/gossip-learn capability (spec §8): decentralized model improvement over any MergeableModel.
    GossipLearner.Builder<M>
    Builder for GossipLearner.
    GossipLearner.Evaluation
    One node's evaluation of one model at the end of an epoch; a metrics-space entry whose authenticated issuer is the evaluating node.
    GossipLearning
    The aspace:cap/gossip-learn capability (spec §8) for parameter-vector models: GossipLearner specialised to double[] under WeightAveraging, so every exchange is a pairwise mean and the fleet's models converge to the fleet mean without any parameter server.
    MergeableModel<M>
    The mergeable-model SPI of aspace:cap/gossip-learn (spec §8): how two peers' models combine on encounter and how a model travels the wire.
    WeightAveraging
    The default MergeableModel: a double[] parameter vector whose merge is the element-wise mean, so a fleet of models converges to the fleet mean and every exchange conserves the fleet sum.