Learning from induced changes in opponent (re)actions in multi-agent games