I feel that a disclaimer is required here: I am not a professionally trained AI engineer. I am self-taught, and most of my knowledge is anecdotal. What follows are my observations that were formed from the creation of the games that I have worked on. Your mileage may vary.
In most games the AI will be given, and allowed to act on, information that it shouldn’t have. In short, it’s allowed to cheat. The AI might be given the location of the player, or knowledge that the player is about to run out of ammunition. Being given information that the AI shouldn’t have isn’t the only way for the game to cheat. The approach called “rubber banding” in gaming AI is the act of tweaking the numbers that drive the AI in response to something that the player does. A game might make an AI’s aim better near the end of a map if the designers want the pressure turned up on the player. You may ask yourself why developers would do this, and the answer is simple: It makes creating a consistent, challenging experience easier. In fact, it may be impossible to make a game consistently challenging and fun without it.
An autonomous AI, on the other hand, can create a completely unique and unpredictable experience every play-through. A player will never know what to expect because the AI isn’t being over-written by the developer stepping in and changing values or behaviors based on what the player is doing. The AI is driving all of it’s own behavior internally. Plus, the developer doesn’t have to give the AI information that it isn’t supposed to have. One of the downsides to the approach, as far as I can see, is that the unpredictable nature of an autonomous AI leads to an inconsistent experience for the player.
When I set out to create Capuchin Capers, it all stemmed from the desire to learn more about AI design. I wanted to attempt to create an AI that would run around the map looking for fruit objects, but without any cheating. At no point should the AI be given the location of a fruit object. In that way, the AI would be autonomous. I added a Director AI to help make the game a bit more challenging for the player when the number of unfound fruit objects remaining reached a certain level. Another reason for the Director AI was because many games use this approach to manage the overall experience that the player is having. I knew in the future that I would want to be able to take advantage of this approach.
I have largely succeeded in my goals. For the most part, the AI will run around the map and pick up fruit objects without any outside aid. And, at times can be quite challenging to defeat. At times. The outcome of the match is heavily dependent upon the placement of the fruit objects. I chose to create a Goal Generator that would randomly place the fruit objects on the ground. In this way the player is never allowed to just memorize where the fruit is. But this also means that the AI won’t always provide a challenge to the player. This is the problem that I am having right now.
Players want a consistent experience. If they choose the easy difficulty for a game, they expect it to be easy…at least comparatively so. If they choose the hard difficulty, they naturally expect it to be much harder than it was when set to easy. A match on easy can’t be more difficult than a different match that was played on the hard setting on the same map. This is inconsistent and frustrating for a player and will probably lead the player to quit playing the game and move on to something else.
A truly autonomous AI is what I wanted to create. An AI that would act on it’s sensory inputs and would make basic decisions based on that information. But in creating this, I have an AI that I can’t easily tweak to provide a consistently challenging experience for the user. I had a goal at the beginning of this project, and that goal was the whole point of the project. I won’t insert code that will allow the AI to cheat. That wasn’t part of the plan, and in my eyes would constitute a complete failure for this project. By adjusting when the Director AI begins providing aid to Suzy (the capuchin monkey), I can adjust the difficulty by some degree. But it won’t provide consistency.
I will do my best to balance the levels so that the AI provides a reasonable experience, even if it is a bit inconsistent. I am sure my AI design is factoring into this issue, so an AI that was designed by someone with more experience would perform much better under the circumstances. I set out to design an autonomous AI that would pick up randomly placed fruit, and that is what I have.
I guess the old adage is true: “Be careful of what you wish for”.