Exploiting locality of interactions using a policy-gradient approach in multiagent learning