LORDRobo

There used to be this BBS door game called Legend of the Red Dragon (L.O.R.D.). Back in early 1994, I wrote a script to play the game for me while I was out smoking or taking a shower. This may have been my first game-playing bot. Some folks call these things "cyborgs" as they play games for the human player over a remote line (in this case, a BBS modem direct connection), substituting for the player's eyes on the monitor and fingers on the keyboard.

At the time that I wrote this, one of my purposes was to make a statement about how cheap Artificial Intelligence (AI) is in most respects. Here was an "intelligent" bot that would

• go to the forest and fight until exhausted,
• collect the treasure and deposit it in the bank,
• purchase healing for wounds between battles,
• exchange gems for increased abilities,
• "talk" at the inn by leaving messages,
• interact with non-player characters (NPCs) that it met in the forest,
• engage in training by combat to increase levels, and
• flirt with the barmaid for points.

LORD was a simple-minded game that, nonetheless, addicted thousands of human players for hundreds of thousands of hours (or more!). The point here is that, under many definitions of "AI", this simple-minded script, basically just a state machine, qualifies as intelligent as it could play this online interactive role-playing game (RPG) just like a human.

Of course, whenever the LORD game was upgraded to a new version, LORDRobo would "break", hanging on a prompt from a new NPC in the forest that it had never met before, unable to relate. This is true of most "intelligent" software: they are brittle. It is my opinion that any definition for "real" AI requires that it be robust and adaptive. To put this another way, AI should be able to deal with novel situations with reasonable responses combined with continuous online learning. Continous online learning ensures that performance can be evaluated and behaviors changed where appropriate. I don't believe that symbolic rule-based AI will ever make it to this level. Instead, I recommend reverse-engineering systems that we know have been successfully delivering intelligence for millions of years: nature's neural networks.