How can a neural network learn to play a game with non-constant number of non-trivial entities?

Question

Suppose you are training a neural network to play a game (like AlphaStar playing StarCraft). The game contains a non-constant number of entities, whose attributes change, and perhaps which have non-constant number of attributes.

For example, using JSON-like notation to illustrate, at time T1 the state might be

state = {
  'health': 90,
  'enemies': [
    {'id': 1, 'x':20, 'y':20, 'items':[]},
    {'id': 2, 'x':200, 'y':200, 'items':[]}
  ]
}

and at time T2 the state might be

state = {
  'health': 90,
  'enemies': [
    {'id': 2, 'x':201, 'y':210, 'items':['coin','bag']},
    {'id': 3, 'x':10, 'y':50, 'items':[]},
    {'id': 4, 'x':50, 'y':10, 'items':[]}
  ]
}

How can time-dependent, variably-sized, and recursively-structured information be given to a fixed-sized neural network?