Welcome to symthic forums! We would love if you'd register!
You don't have to be expert in bit baking, everyone is more than welcome to join our community.

You are not logged in.

Miffyli

Symthic Developer

(6,661)

  • "Miffyli" started this thread

Posts: 3,725

Date of registration
: Mar 21st 2013

Platform: PC

Location: __main__, Finland

Reputation modifier: 17

  • Send private message

1

Friday, March 23rd 2018, 2:50pm

Experimental Self-Learning AI in Battlefield 1

EA/DICE/SEED just published this video: Experimental Self-Learning AI in Battlefield 1 - YouTube

SEED also published a paper on possibly related method they used to train agents in a simpler environment. First glimpses of this were shown months ago + some talk in a conference workshop (iirc): https://arxiv.org/abs/1803.05402
One difference I see is the usage of continuous actions (the aim moves at different speeds and can even "snap").

I do not know about you guys, but this might have made me nerdgasm a little. While they do not seem most efficient AIs, bare in mind they only have the same information human player would have (the screen image and visible numbers)*

I wonder what this will do for cheating: When this technology comes "widely available" (I bet these neural networks can be ran on common gaming machine), will some people use it for "botting"?
Even more interesting is to reach the point where anti-cheat/anti-bot mechanics can not distinguish between human and bot players any more, at which point I we could have said some degree of success in Turing's test.

(*) In case of the linked video this is uncertain. I did not hear narrator say anything about "image only", and the agent seems to lock to enemy through walls.
Links to users' thread list who have made analytical/statistical/mathematical/cool posts on Symthic:
  • 3VerstsNorth - Analysis of game mechanics in BF4 (tickrates, effects of tickrate, etc)
  • leptis - Analysis of shotguns, recoil, recoil control and air drag.
  • Veritable - Scoring of BF4/BF1 firearms in terms of usability, firing and other mechanics.
  • Miffyli - Random statistical analysis of BF4 battlereports/players and kill-distances. (list is cluttered with other threads).
Sorry if your name wasn't on the list, I honestly can't recall all names : ( . Nudge me if you want to be included

NoctyrneSAGA

PvF 2017 Champion

(10,156)

Posts: 7,231

Date of registration
: Apr 3rd 2012

Platform: PC

Battlelog:

Reputation modifier: 19

  • Send private message

2

Friday, March 23rd 2018, 6:31pm

I am also very excited to see learning AIs in Battlefield. Combine good decision-making with unbeatable execution speeds.
Data Browser

Passive Spotting is the future!

"Skill" may indeed be the most magical of words. Chant it well enough and any desire can be yours.

Are you a scrub?

If it flies, it dies™.

Posts: 941

Date of registration
: Dec 14th 2014

Platform: PS3

Location: The Heart of Europe

Battlelog:

Reputation modifier: 8

  • Send private message

3

Friday, March 23rd 2018, 9:37pm

In my diploma thesis I only had the possibility to scratch the surface of Machine Learning with the Neural Networks and developing a response surface, however the results were bad (or better my code was very basic)

The thing is, there must be more than just the neural network to build up.

I tried to receive a response surface from a given data with simulation results and wanted to extrapolate them with a neural network. The results were awful to be honest. I dont want to bash on AI, my knowledge was definitely bad and still is, therefore the bad results.

But I struggle with this knowledge how to implement an AI in that manner, or better how to define a function that "objective" play is rewarded?! How? which input variables, parameters and what's the desired result of the function. How many hidden layers has this thing or are a plethera of NeuralNetworks for different tasks on work here.

I dont want to bash here, just mentioning the questions which come up in my mind from the little knowledge I have.
still playin' Motorstorm

Miffyli

Symthic Developer

(6,661)

  • "Miffyli" started this thread

Posts: 3,725

Date of registration
: Mar 21st 2013

Platform: PC

Location: __main__, Finland

Reputation modifier: 17

  • Send private message

4

Friday, March 23rd 2018, 10:53pm

@ARE5R06

If you are familiar with some of the terminology and you are interested to know the general idea of this side of research (reinforcement learning), I highly suggest you to read the paper they wrote. While not the most novel nor the most detailed, it was rather comfortable to read.

But here's a tl;dr version: You give agent cookies when it does good (e.g. kills enemies), and agent attempts to learn a behavior which results in as many cookies as possible (e.g. most kills)
Links to users' thread list who have made analytical/statistical/mathematical/cool posts on Symthic:
  • 3VerstsNorth - Analysis of game mechanics in BF4 (tickrates, effects of tickrate, etc)
  • leptis - Analysis of shotguns, recoil, recoil control and air drag.
  • Veritable - Scoring of BF4/BF1 firearms in terms of usability, firing and other mechanics.
  • Miffyli - Random statistical analysis of BF4 battlereports/players and kill-distances. (list is cluttered with other threads).
Sorry if your name wasn't on the list, I honestly can't recall all names : ( . Nudge me if you want to be included

NoctyrneSAGA

PvF 2017 Champion

(10,156)

Posts: 7,231

Date of registration
: Apr 3rd 2012

Platform: PC

Battlelog:

Reputation modifier: 19

  • Send private message

5

Saturday, March 24th 2018, 12:02am

AI right now can be described as a search algorithm looking for the best move in a giant space of possible moves.

As my professor once said, "Even though we call it AI, the intelligence part still isn't there yet."
Data Browser

Passive Spotting is the future!

"Skill" may indeed be the most magical of words. Chant it well enough and any desire can be yours.

Are you a scrub?

If it flies, it dies™.

Posts: 548

Date of registration
: Dec 24th 2011

Platform: PC

Battlelog:

Reputation modifier: 10

  • Send private message

6

Saturday, March 24th 2018, 9:28am

This is really some complex, yet fascinating, stuff.

But still, I'm a simple man. Add an offline bot mode, and I'll be happy. Perhaps I would possibly like to return to Battlefield 1 in like ten years, when the servers are almost-certainly shut down, and actually just piss around with some of the fun toys and (then to be) outdated mechanics that I would remember. The old Refractor Battlefield titles are still a tiny bit fun to mess around with since you don't actually need to be in an online match with actual people to play them.

Miffyli

Symthic Developer

(6,661)

  • "Miffyli" started this thread

Posts: 3,725

Date of registration
: Mar 21st 2013

Platform: PC

Location: __main__, Finland

Reputation modifier: 17

  • Send private message

7

Saturday, March 24th 2018, 12:01pm

But still, I'm a simple man. Add an offline bot mode, and I'll be happy. Perhaps I would possibly like to return to Battlefield 1 in like ten years, when the servers are almost-certainly shut down, and actually just piss around with some of the fun toys and (then to be) outdated mechanics that I would remember. The old Refractor Battlefield titles are still a tiny bit fun to mess around with since you don't actually need to be in an online match with actual people to play them.


I share this same wish! It would be cool to experience those pre-internet times (for me) by playing against mute but "intelligent" enough bots ^^. They could even tune them to "roleplay" the game, so it would be a proper BF1 multiplayer round but more of the "authentic" taste. I think this should be technically feasible too, given the fine-tuning game developers would do for the AI code before putting it out, so it should be runnable on player's computer.
Links to users' thread list who have made analytical/statistical/mathematical/cool posts on Symthic:
  • 3VerstsNorth - Analysis of game mechanics in BF4 (tickrates, effects of tickrate, etc)
  • leptis - Analysis of shotguns, recoil, recoil control and air drag.
  • Veritable - Scoring of BF4/BF1 firearms in terms of usability, firing and other mechanics.
  • Miffyli - Random statistical analysis of BF4 battlereports/players and kill-distances. (list is cluttered with other threads).
Sorry if your name wasn't on the list, I honestly can't recall all names : ( . Nudge me if you want to be included

Posts: 1,227

Date of registration
: Dec 7th 2011

Platform: PC

Battlelog:

Reputation modifier: 12

  • Send private message

8

Saturday, March 24th 2018, 1:37pm

I wonder how well these bots would do against human players in their current state?
bob

Miffyli

Symthic Developer

(6,661)

  • "Miffyli" started this thread

Posts: 3,725

Date of registration
: Mar 21st 2013

Platform: PC

Location: __main__, Finland

Reputation modifier: 17

  • Send private message

9

Saturday, March 24th 2018, 2:04pm

I wonder how well these bots would do against human players in their current state?


Based on interview here:

Quoted from "Magnus Nordin - SEED Tech. Director"

We have conducted playtests, pitting AI agents against human players in a simplified game mode, restricted to handguns. While the human players outperformed the agents, it wasn’t a complete blowout by any stretch.


And judging by the video, the agents are do realize how to fire at the enemy and aim for health/ammopacks when needed. They mention "bots lack planning", and if the design of the method has not changed much since the paper, the bots do not have long term memory of anything. They also seem to be limited to primary (and/or secondary?) weapon. It is also common to resize the screen image to something relatively small (128x128 in the paper), so the agents are not able to see very far (these are patch-arounds for this, but it is hard to say if they use these).

I would hazard a guess they are "decent" in TDM and DOM games. In anything larger / more complicated they likely get stuck or are uncertain what to do.

This seems similar to results in the VizDoom competitions (DOOM AIs fighting each other):
- In the track1 of the competition: Only one (known) map and players only used rocket launchers. Bots are quite decent in this one, and they do give some challenge to human players (subjective opinion: The map is a clusterbuck and mindless shooting :'D).
- In the track2 of the competition: Multiple unknown maps, weapons had to be picked up from level (i.e. normal DOOM deathmatch). They are still able to fight each other, but seem to be stumbling around randomly and lose track of enemies at times. They do not seem to have long-term planning, and easily forget they have been spinning around in the same location for quite a while (certified goldfish-grade memory?)
Links to users' thread list who have made analytical/statistical/mathematical/cool posts on Symthic:
  • 3VerstsNorth - Analysis of game mechanics in BF4 (tickrates, effects of tickrate, etc)
  • leptis - Analysis of shotguns, recoil, recoil control and air drag.
  • Veritable - Scoring of BF4/BF1 firearms in terms of usability, firing and other mechanics.
  • Miffyli - Random statistical analysis of BF4 battlereports/players and kill-distances. (list is cluttered with other threads).
Sorry if your name wasn't on the list, I honestly can't recall all names : ( . Nudge me if you want to be included