Improving learning in robot teams through personality assignment

This paper presents two reinforcement learning algorithms, which are inspired by human team dynamics, for autonomous robotic agent applications. These algorithms entail strictly local credit assignments to individual agents and hence promote team scalability. The first algorithm is termed the Golden Rule Learner (GRL) and incorporates agent self-reward for completion of altruistic actions in addition to self-reward when completing team goals. The second algorithm is termed the Personality Adjusted Learner (PAL), which extends the GRL algorithm by using the human-oriented Myers–Briggs Type Indicator (MBTI) as the inspiration for assigning weights to the local rewards earned by PAL agents, simulating individual agent personality preferences for types of tasks. In this way, it contributes to the wider research goal of creating a real-life computational equivalent of the human mind by providing a mathematical mechanism for encoding and processing personality preferences that can be described by systems such as the MBTI, and are a critical aspect of human interaction. The work presented in this paper tests the hypothesis that the assignment of MBTI preferences improves the agents’ team performance over the baseline and GRL teams for a given task.A resource gathering scenario was simulated using teams of agents. These scenarios include: a baseline team of nonaltruistic locally rewarded agents, GRL agents, or a team of PALs to gather resources. When the resources were scarce, at least one possible combination of PAL personalities showed superior performance over the baseline and GRL agents. This is an indicator that the MBTI specification of PAL agents can be used to optimize team performance.