[Discussion] Human I vs. AI - Risk tolerance, Math, and irrationality
Elemental Forums
"Evaluate" - notice this word comes from the root word, Value. It means to attach numerical values for the purpose of comparison and decision making.
Intelligence, as it relates to games and game theory, has basically two components - Evaluation and Temporal awareness. You must be able to make a decision on some sort of sensical basis. You must be able to differentiate between "now" and "not now". The first component is the dominant one.
For humans, we call the latter bit "learning". We remember what happened in the past, and that helps in decision making in the present, as well as anticipating what will happen in the future (which also helps in decision making in the present).
For Artificial awareness, this learning element is difficult to approximate. It requires independent thought, or true randomness - which a computer can't do. Your calculator doesn't just do math all on it's own - it requires input to generate output.
Instead, the "past" learning element is essentially encoded into an artificial intelligence, in the form of attached values to possible events and decisions. These are not generated at random from experiential data, although advanced methods will update and modify values as a result of experience. The "future' aspect is simply a matter of looking X 'moves' ahead, and looking at the desirability of various possible outcomes, weighted by probability (which is often encoded as well). This is, incidentally, how many computer chess programs function, with the 'difficulty' capping the number of moves-forward and possible outcomes the computer will evaluate.
Once you have a value, it's just a matter of Goal - something like maximize, minimize, or approach/stay within range. This is, on some level, true of all decision making, whether artificial or not.
For example, let's say you attach a value of 5 to playing Elemental (arbitrary). You attach a value of 3 to sleeping, and a value of 4 to eating. If your goal is to maximize value, You'll just keep playing elemental and never eat or sleep. But let's say Elemental's AI or crashes frustrates you generating a value of -2. The result is, you'll play Elemental (+5), get pissed (5-2 = 3) and then go eat... or play more elemental anyway.
Human opponents inject the issue of randomness - people attach different values to different things, generating different choices. It can be unpredictable, and quickly change. This is because human intellect (or lack thereof) isn't bound by strictly circumscribed rules - see Cognitive Dissonance for more. Rather than using purely numerical and analytical decision trees (value of outcome weighted by probability of outcome to generate average/"true" value), humans often skip the math - an irrationality element kicks in and overrides values.
This is also complicated by 'risk tolerance.' The degree to which discount or inflate the value of a risky outcome - in short, how much risk we are willing to tolerate or endure. Some people are highly risk-seeking: we call them base jumpers. Some people are highly risk intolerant: we call them insurance adjusters.
Take for example, my most recent scenario in Elemental:
Player is bounded on all sides by hostile factions. Two narrow mountain passes to the north lead to a militarily advanced Krax empire. A broad forest region to the South west leads to a resource rich Altar kingdom. Player heavily favors AoE magic attacks, sovereign and imbued champion led armies of summons, and disfavors heavily protecting settlements. Player controls one Air and one Water shard.
Player sends an army lead by the Queen (spouse) to attack Altar because altar controls an Air shard. The altar settlement is lightly guarded. The Queen's unit: Champion, Vigilant minion, Lord of Ice, Stone Giant, Fire giant, etc. etc. Because of the way movement is handled, the Queen's unit proceeds one tile at a time towards the Altar settlement, after dispatching two intervening pioneers. It will reach the settlement in two turns.
Now, it's not totally clear what the altar AI was doing - but apparently it was moving it's sovereign Relias very slowly towards the settlement, along with a large army. It didn't teleport - and I don't know why. Nor did it train any units or further increase the defense of what turned out to be the initial settlement (spouse and children present).
The settlement was crushed and became player territory. In the immediate next turn, Altar's sovereign and army attacks (despite having a lower combat rating). They are soundly defeated, and because the sovereign died in enemy lands, Altar's entire faction was eliminated.
What?
If I had been playing a human opponent, it might be possible to explain this outcome. Altar may have attached an inordinately high probability to it's success. It may have attached an inordinately high value to it's initial settlement. It may have rationally decided that having a foreign power bifurcate it's territory was highly undesirable. It may have irrationally desired revenge. It may have saw a very high value in reducing my military by one third, despite a low probability of success. The human player may have been exceptionally risk seeking - either because it saw itself as defeated, so any degree of risk was acceptable, or simply a "gambling man". A more risk-avoiding player would have written settlement off as lost, and amassed a counter strike.
But for a strictly mathematical AI (1 is always less than 2, unlike with meatbags), it didn't make any sense. In short, the AI had no conception whatsoever of risk or future events. The possibility of losing the sovereign and losing it's entire faction (complete loss) didn't enter into it's evaluation at all. It was very simply, something on the order of, "I was attacked, therefore attack back."
What we might have expected is that the AI would have considered the disparate combat ratings (probability of loss better than average) and the highly unfavored outcome of complete loss to generate an action of either "do not attack" or (for a more complicated system) "send sovereign to safe haven, attack with remaining forces".
A far more complicated AI system (more moves, more tables of values, and considers opponent's table of values) would probably have employed a "scorched earth" tactic - burn down the air shard to make my victory less meaningful (this actually happened in another game, but purely by accident - the AI had a caravan path through the shard temple, so on the subsequent turn, caravans automatically rolled right through and permanently destroyed the shard).
Understanding all this, it's pretty easy to see why every AI faction plays almost identically (aggressive, early expansion, followed emphasis on large numbers of grouped army units - incidentally, I have no idea how it gets the gildar and resources to support that...), why every tactical battle proceeds in almost the same manner (hit end turn twice, smite foes; enemy always targets familiar, etc. first)... and why there is a lack of variety (every unit moves 2).
The underlying AI - the table of values attached to units, events, decisions, etc. and the decision making process/goal - is extremely basic. Consequently, the values it needs to consider are also very simple (and often heterogeneous, allowing for simple comparisons and decisions).
The upshot is, the underlying AI needs to be heavily upgraded... and that would take a while, since developers/game theorists need to consider a wider range of events and potential player strategies to 'teach' the AI. And that's IF there were no changes being introduced. With the introduction of various other changes (tweaking weapons damage, magic system, etc.) expect the current AI to make some bizarre decisions. It could be a while before it becomes challenging in a sensical manner - as well as incorporate features we're all screaming for (for example 'personality' - like an AI that favors a tactic of 'mine or no-ones' by scorching resources and npcs).
