[HN Gopher] How to record data for reinforcement learning agent ...
___________________________________________________________________
How to record data for reinforcement learning agent from any Linux
game (2020)
Author : mooreds
Score : 50 points
Date : 2022-03-11 02:10 UTC (3 days ago)
(HTM) web link (medium.com)
(TXT) w3m dump (medium.com)
| alfanick wrote:
| The title is misleading, it promises recording data from *any*
| game, yet it requires some UDP telemetry to be implemented in the
| game itself and uses F1 as the game.
|
| I was expecting something with a) strace to track IO or b) gdb to
| monitor specific memory regions which contain game state or c)
| some eBPF stuff (to track IO or memory aka a+b combined).
| quanto wrote:
| The method discussed in OP requires UDP telemetry to be
| implemented by the game. Not every game will have this option.
|
| I recall a researcher friend who tried RL on an arbitrary game of
| his choice by exploiting the system memory used by the game and
| using the memory state as a game state. Of course, the memory
| state is not human-readable (unless you have the source code and
| the intimate knowledge of the game; even so, it would be a
| difficult thing to do by a human) but a clever RL algorithm would
| treat the state just the same as from a human-readable state
| space.
|
| His project was a wild success and beat an experienced human
| player. I recall he had a youtube video somewhere. I miss that
| guy.
| CaptainNegative wrote:
| Dumb clickbait title. Only works for games with built-in "UDP
| telemetry".
|
| Next up: how to find all solutions to any 3SAT instance in
| polynomial time. Just read from the accompanying
| 3sat_instance_solutions.txt .
___________________________________________________________________
(page generated 2022-03-14 23:02 UTC)