I’ve been pondering a game; an iterated prisoner’s dilemma with extended rules revolving around trading information.
Utility points can be used between rounds for one of several purposes; sending messages to other agents in the game, reproducing, storing information (information is cheap to store, but must be re-stored every round), hacking, and securing against hacking.
There are two levels of iteration; round iteration and game iteration. A hacked agent hands over its source code to the hacker; if the hacker uses its utility to store this information until the end of the current game, the hacking agent’s programmer gets the source code of the hacked agent. Thus, the programmer can reformulate his agents to target that source code between games.
Utility and agents are carried over between games; the programmers can expend their utility between games to reprogram their agents (a fixed cost per agents they have). They can also expend utility to “Defend world infrastructure”, which increases the cost of the endgame action (initially already prohibitively expensive, to require several game iterations), “Seize control of world infrastructure.”
Round actions can only be performed against known agents; an agent becomes known and identifiable by playing a round of prisoner’s dilemma with them, or by getting their ID from communication. Agents can lie; they can misidentify another agent as cooperating when in fact it routinely defects. They can also transmit arbitrary information; all agents of a given player could use a handshake in order to identify fellow agents of that player. (Which could introduce a greater vulnerability to hacking, as each agent must defend against hacking individually.)
An agent who runs out of utility dies.
My question is, does anyone think a contest between unfriendly versus friendly AI be effectively modeled in this game? If so, would it be better modeled as a free-for-all, a single friendly “avatar” fighting a sea of unfriendly AI, or a single unfriendly “avatar” fighting a sea of friendly AI? Would adding “friendly versus unfriendly” as the game’s target of modeling require some additional available actions to reflect the underlying purpose of the agents in the game?
Disregarding the question of actual AIs¹, this sounds like it would make for an awesome browser-based “hacking” strategy game. It could also fit well into a game design similar to Uplink) or Street Hacker.
¹. (I’m not good enough with AI theory yet to really have any useful insight there)
I’ve been pondering a game; an iterated prisoner’s dilemma with extended rules revolving around trading information.
Utility points can be used between rounds for one of several purposes; sending messages to other agents in the game, reproducing, storing information (information is cheap to store, but must be re-stored every round), hacking, and securing against hacking.
There are two levels of iteration; round iteration and game iteration. A hacked agent hands over its source code to the hacker; if the hacker uses its utility to store this information until the end of the current game, the hacking agent’s programmer gets the source code of the hacked agent. Thus, the programmer can reformulate his agents to target that source code between games.
Utility and agents are carried over between games; the programmers can expend their utility between games to reprogram their agents (a fixed cost per agents they have). They can also expend utility to “Defend world infrastructure”, which increases the cost of the endgame action (initially already prohibitively expensive, to require several game iterations), “Seize control of world infrastructure.”
Round actions can only be performed against known agents; an agent becomes known and identifiable by playing a round of prisoner’s dilemma with them, or by getting their ID from communication. Agents can lie; they can misidentify another agent as cooperating when in fact it routinely defects. They can also transmit arbitrary information; all agents of a given player could use a handshake in order to identify fellow agents of that player. (Which could introduce a greater vulnerability to hacking, as each agent must defend against hacking individually.)
An agent who runs out of utility dies.
My question is, does anyone think a contest between unfriendly versus friendly AI be effectively modeled in this game? If so, would it be better modeled as a free-for-all, a single friendly “avatar” fighting a sea of unfriendly AI, or a single unfriendly “avatar” fighting a sea of friendly AI? Would adding “friendly versus unfriendly” as the game’s target of modeling require some additional available actions to reflect the underlying purpose of the agents in the game?
Disregarding the question of actual AIs¹, this sounds like it would make for an awesome browser-based “hacking” strategy game. It could also fit well into a game design similar to Uplink) or Street Hacker.
¹. (I’m not good enough with AI theory yet to really have any useful insight there)