Implicit estimation of other´s intention without direct observation of actions in a collaborative task: situation-sensitive reinforcement learning

An agent in a multi-agent environment should adapt to the diversities of dynamics that are caused by changes in the physical properties of the task environment and in social situations concerning how the partner is shifting his/her behaviors to achieve the task. When the partner´s intention changes in the latter, a collaborator agent has to notice this from what is observed in the shared-task environment and to explore how to adaptively collaborate with the partner. A situation-sensitive reinforcement learning (SSRL) architecture is presented in this paper. SSRL enables a collaborator agent to implicitly estimate the partner´s goal. The mathematical basis of the implicit estimates is also addressed. A simple truck-pushing task by a pair of agents is presented as a testbed example, and the simulation results show that organized collaboration could be achieved by an agent embedded with our model in adapting to the partner´s intentional strategic changes.