Abstract: A computer accesses a dataset for computing an action including multiple decision items and a configuration for a directed acyclic graph (DAG) comprising nodes representing dimensions. Each dimension is associated with a decision item of the action. The computer computes a value for a first decision item associated with a top level dimension of the DAG based on the dataset and using a reinforcement learning agent for the top level dimension. The computer computes a value for a second decision item associated with a non-top level dimension from the DAG based on the dataset and a specified subset of outputs of reinforcement learning engines for dimensions upstream from the non-top level dimension in the DAG and using a reinforcement learning agent for the non-top level dimension. The computer provides the computed action including the value for the first decision item and the value for the second decision item.

Abstract: A computing machine receives a request for an action by a meta-agent at the computing machine. The computing machine selects, using the meta-agent, an agent for determining an action responsive to the request. The agent is selected from a plurality of reinforcement learning agents. The agent is selected based on values in a dataset associated with the request. At least a portion of the dataset has predefined value types. The meta-agent and each agent from the plurality of reinforcement learning agents is trained using datasets including the predefined value types. The computing machine transmits, to the selected agent, the request for the action to cause the agent to compute the action.