simai

High-level TODOs for AI simulation code:

• When copying the game and simulating, copy what's on the stack. This will allow AI to use effects that e.g. sacrifice their creatures that are targeted by removal as well as play counterspells using the simulation framework.

• When evaluating permanents with ETB effects (e.g. oblivion ring), simulate all possible targets, just like SA targeting during simulation.

• When evaluating what to play during combat, simulate until end of combat, so that the AI takes into account how much life or chump blockers are saved when killing an attacking creature, for example.

• When evaluating game state, creature evaluations should take into account the current turn and phase and what creatures the opponent has - so that for e.g. if the opponent can kill you next turn with enough attacking creatures, then effects that produce blockers are rated higher.

• AI is too eager to play spells. For example, the AI may want to play an instant during draw phase, since that's the best spell to play at that point, whereas in reality, waiting till MAIN1 and playing a creature or sorcery may be the better decision.

• AI needs to ration mana somehow. Right now, it's a "greedy play best spell" system. It fails to take into account situations where playing two cheap spells is a better alternative than playing one expensive spell whose effect is better than each cheap spell, but not than both of them.

Ideally we'd work towards a smoother integration between simulation and heuristics, e.g. while the human thinks on his priority the simulation can already start evaluating and if the human passes without a change in board state the plan can be reused (otherwise fall back to heuristics after some defined threshold)