A recent arXiv preprint tackles a known weakness of deep reinforcement learning agents in real-time strategy games: they can be strong against familiar opponents yet fail against those outside their training distribution. The authors propose a constrained, command-conditioned reinforcement learning framework that separates high-level strategic command selection from low-level learned unit control.
Instead of letting a single policy handle both strategy and tactics, the method uses a bandit strategy selection mechanism to choose among different strategic options. This separation is intended to make the agent more flexible when facing novel opponent behavior, since the strategic layer can adapt without retraining the entire control policy.
The abstract is brief and does not include experimental results, so the effectiveness of the approach is not yet demonstrated in the available text. The contribution at this stage is the architectural idea: decoupling strategy from control and using bandit selection to handle distribution shift in competitive settings.