So you are saying that if you first trained it on a different game with similar real-world mechanics, then training it on this game would require far fewer iterations?

Pac-Man isn't a purely abstract game, it's loaded with human cultural references. Pac-Man himself is a creature who loves eating food: mostly white pellets but also various kinds of fruit and of course the iconic power pellets. Fruit of course is something humans immediately identify as something good to eat and in the game they're rewarded with points which are denoted by numbers on the screen that grow larger.

It almost needn't be said that human culture puts a great deal of emphasis on making numbers grow larger, particularly those associated with personal wealth and economic growth. To a computer without this context, why should there be any reason to prefer larger numbers over smaller ones? Heck, if you're going all the way and denying the AI direct access to the game's memory, then it's got to learn to recognize the digits first before it can even begin to decide whether it's good to increase them.

Even more abstractly, the game relies on the human concepts of good and bad. It's good to eat the fruit and pellets to increase your score counter. It's bad to let a ghost catch pac-man, causing him to die and lose a life. Without being taught, human children can identify all of these elements but the computer would seem to have a more difficult task.

This called transfer learning and is an active area of research. It doesn't work nearly as well as it "should" if "AI" was really Artificial "Human Intelligence".

It might. My understanding is, it is common to take trained models from one problem and use parts of them on other problems in a similar domain to speed up training.

No, in humans the cycles involved in walking and 3d orienting carry over, so several thousand iterations of pre-game and a few of game.

This is a tired argument, experiments on humans with random textures show only a 10X slowdown, not 10,000x

Indeed. Some work has actually been done to try to quantify how important various types of human prior knowledge are for learning video games: https://arxiv.org/pdf/1802.10217.pdf

Humans intelligence basically evolved (as far as we know) by retarding childhood development so we had neural plasticity longer to learn more. Crazy that we're giving our already-dumb AIs the extra challenge of learning things from nothing and doing so quickly.