0
$\begingroup$

I was discussing with a friend whether current AI does anything remotely similar to 'thinking' and he argued that AIs that play games must think up strategies.

While thinking may not be precisely defined, my understanding of algorithms like OpenAI was that they just minimize a very non-convex objective, but still play the game based on examples, and not by coming up with intentional strategies. Is my understanding incorrect?

$\endgroup$
  • $\begingroup$ Advanced gameplaying agents like “OpenAI five” aren't searching for a policy in the state space which optimizes the outcome of the game but they are using a more advanced technique. Instead of creating a strategy (=sequence of steps to win a game), modern game playing agents are using a combination of macro actions, symbolic forward simulation and managers for subgames (e.g. for determine the build order). $\endgroup$ – Manuel Rodriguez Nov 13 '19 at 13:16
1
$\begingroup$

The key here is think up strategies. If we define this as examining, creating a hypothesis, and testing it as strategizing then yes AI has the ability to strategize. It can examine other users' games, quantifies actions that correlated with victory then test if it gains victory by doing those actions.

Strategy by definition is: a plan of action or policy designed to achieve a major or overall aim.

AI can not classically plan a series of actions designed to achieve a major victory. Instead, it learns the right strategy by testing simulated scenarios, like someone who thinks about the consequences before doing them, but the AI actually has the opportunity to play the game hundreds of times, in order to learn the correct strategy. Similar to Bill Murray in the movie Groundhog Day, learning the ideal day to live. The AI can strategize by experiencing the game over and over until it fine-tunes what an ideal game should be and has seen enough examples of games to not be outwitted by gimmicky strategies.

To summarize, AI can strategize, just in a way fairly different than people.

$\endgroup$
  • $\begingroup$ Thanks you for your insight. I guess my point of contention to 'thinking up strategies' can then better be described as: The AI does not come up with a strategy by means of contemplation, but by randomly trying (simulating scenarios) and choosing only the winning versions... Right? $\endgroup$ – Frans Rodenburg Nov 14 '19 at 2:00
  • $\begingroup$ Yes, although a bot could technically also train on pushing the game toward states that exist in wins and variables in the game associated with winning too. So the bot examines and assimilates win data and meta win data as well. Maybe during part of a lost game the team was winning then incorrect actions caused a loss, a good bot like a good person should be able to ‘learn’ from loss games and win games because not every decision or action in a lost game is incorrect. $\endgroup$ – Michael Hearn Nov 14 '19 at 2:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.