There is evolving notion of stacked reinforcement learning systems, e.g. https://www.ijcai.org/proceedings/2018/0103.pdf - where one RL systems executes actions of the second RL system and it itself executes action of the thrid and the reward and value flows back.
So, one can consider the RL system RLn with:
- S - space of states;
- A={Ai, Ainner, Ao} - space of actions which can be divided into 3 subspaces:
- Ai - actions that RLn exposes to other RL systems (like RLn-1, RLn-2, etc.) and for whome the RLn provides reward to those other systems;
- Ainner - actions that RLn executes internally under its own will and life;
- Ao - actions that RLn imports from other RL systems (like RLn+1, RLn+2, etc.) and that RLn executes on the other systems and form which RLn gets some rewards, that can be used for providing rewards for Ai actions.
So, one can consider the network (possibly hierarchical) of RL systems . My questions are:
- is there some literature that consider such stacks/networks of reinforcement learning systems/environments?
- is there economic research about value flow, about accumulation of wealth, about starvation and survival proceses and evolution in such stacks/networks of RL systems/environments
Essentially - my question is about RL environments that can function in the role of agents for some other environments and about agents that can function in the role of environemnts for some other agents. How computer science/machine learning/artificial intelligence name such systems, what methods are used for research and how the concepts of economics/governance/preference/utility are used in the research of evolution of such systems?
