抄録
A learning automaton is connected in feedback loop to the random environment, where the input to one side is the output to the other side. To use such abstract machine for formulating reinforcement learning systems, we need to generalize it into some new situations, which can correspond to context inputs from environment.
Here, we propose a new type learning automaton with context inputs on multi-teacher environments. For formulating some tradeoff environments, we present two descriptions, weighted majority decision and accumulated relative error decision.
We also present a strong theorem for proposed reinforcement scheme operating under some conditions. The simulation results show that, on static tradeoff environments, our reinforcement scheme has better properties in convergence speed and shaped learning processes.
