A new reinforcement learning method lets AI agent teams reshuffle on the fly without retraining.
Researchers introduce PLATO, short for Pointer Learner for Agent and Task Openness, which pairs a pointer-network actor with a graph neural network critic, trained using multi-agent proximal policy optimization. It targets what the paper calls open agent systems, where the number of agents and the set of tasks can both change unpredictably, a scenario standard multi-agent reinforcement learning handles badly because it assumes fixed state and action spaces. Current workarounds pad and mask the action space to fake a fixed size, or use graph-based methods that only cope with one kind of change at a time. PLATO's actor instead outputs a probability distribution directly over whatever tasks currently exist, while its critic encodes agents and tasks as a graph that grows or shrinks along with them. The team formalized this setup mathematically and tested it in a wildfire suppression simulation from the MOASEI benchmark, where PLATO showed more consistent zero-shot generalization than existing baselines.
Most real coordination problems, wildfire response, delivery fleets, disaster relief, do not come with a fixed roster of agents or a fixed task list. A method that only works at a set team size is a lab exercise, not a deployable tool. Padding and masking are patches on that limitation, not a fix, since they still cap how large a system can grow before retraining is needed.
One wildfire simulation is not a fleet of delivery drones or a hospital staffing system, so whether PLATO's trick holds up outside a controlled benchmark is still an open question.