• Log InLog In
  • Register
Liquid`
Team Liquid Liquipedia
EDT 17:38
CEST 23:38
KST 06:38
  • Home
  • Forum
  • Calendar
  • Streams
  • Liquipedia
  • Features
  • Store
  • EPT
  • TL+
  • StarCraft 2
  • Brood War
  • Smash
  • Heroes
  • Counter-Strike
  • Overwatch
  • Liquibet
  • Fantasy StarCraft
  • TLPD
  • StarCraft 2
  • Brood War
  • Blogs
Forum Sidebar
Events/Features
News
Featured News
Maestros of the Game: Week 1/Play-in Preview12[ASL20] Ro24 Preview Pt2: Take-Off7[ASL20] Ro24 Preview Pt1: Runway132v2 & SC: Evo Complete: Weekend Double Feature4Team Liquid Map Contest #21 - Presented by Monster Energy13
Community News
LiuLi Cup - September 2025 Tournaments2Weekly Cups (August 25-31): Clem's Last Straw?39Weekly Cups (Aug 18-24): herO dethrones MaxPax6Maestros of The Game—$20k event w/ live finals in Paris49Weekly Cups (Aug 11-17): MaxPax triples again!15
StarCraft 2
General
Production Quality - Maestros of the Game Vs RSL 2 Heaven's Balance Suggestions (roast me) Maestros of the Game: Week 1/Play-in Preview Team Liquid Map Contest #21 - Presented by Monster Energy Weekly Cups (August 25-31): Clem's Last Straw?
Tourneys
Maestros of The Game—$20k event w/ live finals in Paris RSL: Revival, a new crowdfunded tournament series LiuLi Cup - September 2025 Tournaments Sea Duckling Open (Global, Bronze-Diamond) Sparkling Tuna Cup - Weekly Open Tournament
Strategy
Custom Maps
External Content
Mutation # 489 Bannable Offense Mutation # 488 What Goes Around Mutation # 487 Think Fast Mutation # 486 Watch the Skies
Brood War
General
ASL20 General Discussion BGH Auto Balance -> http://bghmmr.eu/ BW General Discussion Pros React To: herO's Baffling Game BSL Polish World Championship 2025 20-21 September
Tourneys
[Megathread] Daily Proleagues Is there English video for group selection for ASL Small VOD Thread 2.0 [ASL20] Ro24 Group F
Strategy
Simple Questions, Simple Answers Muta micro map competition Fighting Spirit mining rates [G] Mineral Boosting
Other Games
General Games
General RTS Discussion Thread Stormgate/Frost Giant Megathread Nintendo Switch Thread Path of Exile Warcraft III: The Frozen Throne
Dota 2
Official 'what is Dota anymore' discussion
League of Legends
Heroes of the Storm
Simple Questions, Simple Answers Heroes of the Storm 2.0
Hearthstone
Heroes of StarCraft mini-set
TL Mafia
TL Mafia Community Thread Vanilla Mini Mafia
Community
General
US Politics Mega-thread Things Aren’t Peaceful in Palestine Russo-Ukrainian War Thread Canadian Politics Mega-thread YouTube Thread
Fan Clubs
The Happy Fan Club!
Media & Entertainment
[Manga] One Piece Anime Discussion Thread Movie Discussion! [\m/] Heavy Metal Thread
Sports
2024 - 2026 Football Thread Formula 1 Discussion TeamLiquid Health and Fitness Initiative For 2023
World Cup 2022
Tech Support
Computer Build, Upgrade & Buying Resource Thread High temperatures on bridge(s)
TL Community
The Automated Ban List TeamLiquid Team Shirt On Sale
Blogs
Collective Intelligence: Tea…
TrAiDoS
A very expensive lesson on ma…
Garnet
hello world
radishsoup
Lemme tell you a thing o…
JoinTheRain
RTS Design in Hypercoven
a11
Evil Gacha Games and the…
ffswowsucks
INDEPENDIENTE LA CTM
XenOsky
Customize Sidebar...

Website Feedback

Closed Threads



Active: 1097 users

Towards a good SC bot - P5 - H.I. (1/2)

Blogs > imp42
Post a Reply
imp42
Profile Blog Joined November 2010
398 Posts
Last Edited: 2016-11-23 15:40:40
November 23 2016 04:57 GMT
#1
Towards a good StarCraft bot - Part 5 - "Human Intelligence (1/2)"

Summary:
+ Show Spoiler +
A short overview over the executive functions (cognitive capabilities) I consider most relevant for StarCraft players - human or artificial.


While Neural Nets keep me busy, it is a good time for a change of perspective. If we want good Artificial Intelligence bots, what can we copy from Human Intelligence? How do humans work?

That is obviously a very broad question. It involves many different subfields. Here, I pick some aspects that I personally consider important and related to the ability to play games like StarCraft on a high level.

What is human intelligence?
There are many different definitions. So, I just pick one that suits my purpose: human intelligence is the intellectual capacity characterized by perception, consciousness, self-awareness, and volition. It gives humans the cognitive abilities to learn, form concepts, understand, and reason.

Can intelligence be measured?
As the definition of intelligence suggests, there are several dimensions to intelligence. However, there is a notion of general intelligence called “g” or “g factor”. Supporters of this notion believe “g” to be a good metric because it summarizes performance among different cognitive tasks. Many different tests have been developed around such cognitive tasks, to capture the dimensions of intelligence. One test that is believed to be positioned at the center of all those dimensions is called the Raven Progressive Matrices test. This shown in the following illustration:

[image loading]

Figure taken from: Carpenter, P. A., Just, M. A., & Shell, P. (1990). What one intelligence test measures: a theoretical account of the processing in the Raven Progressive Matrices Test. Psychological Review, 97(3), 404–431.

Executive functions
As we learned, intelligence gives us the cognitive abilities required to e.g. play a game of StarCraft. There are ways to measure intelligence and tests have been developed in an attempt to do so. Many tests attempt to capture dimensions of intelligence by measuring performance of a set of execute functions. Executive functions are a set of cognitive processes highly relevant to how well we play StarCraft*. Because these cognitive processes include attentional control, working memory, reasoning, problem solving, and (very important) planning.

*+ Show Spoiler +
On a side note, the same processes are also highly relevant in assessments to identify high-potentials and future leaders in corporates. But we don’t really care about that, right?


Let me elaborate a bit and put each of the mentioned processes into the context of playing a game of StarCraft:

Attentional control: the capacity to choose what to pay attention to and what to ignore. Clearly relevant when perceiving all the graphical (and audio) input from the game. Indicators of mineral/gas income, max supply/used supply, health bars, moving dots on the mini-map, etc. etc. This also includes the ability not to pay attention to the nice reflections in the water or the cute animation of the critter.

Working memory: the system responsible for holding, processing, and manipulating information. There are only so many things we can keep in our working memory at once. A rough estimation is 5 +/- 2 (i.e. between 3 and 7). This implies even the best player is prone to make mistakes when he needs to deal with more than 7 issues simultaneously. Cleary StarCraft overburdens human working memory, so we need to be selective (see attentional control).

Reasoning: the capacity to apply logic and establish facts and beliefs. “If we are at the t minute mark in the game and the opponent already has x units of type y out, then he can impossibly have any units of type z” is an example of such reasoning. Or more simply: “My scout arrives at his base and the base is empty -> he must have a proxy somewhere”.

Problem solving: The capacity to use generic or ad hoc methods to find solutions to problems. Let’s say you played ladder some years ago and ran into a player called “Gaulzi”. You would immediately know he _only_ does cannon rushes. And sure enough, you find a pylon warping in behind your mineral line. Since you already knew it was coming you can apply a generic “defend against canon rushes” method, even though you may have never faced the exact same situation (same map, same spots, pylon on the same tile, etc.) before. Let’s now assume you ran into TLO in the early days. You might end up playing against something you have never seen before. Your ability to come up with an answer on the spot is referred to as ad-hoc problem solving.

Planning: the capacity to think about and organize activities required to achieve a required goal. In StarCraft it is very beneficial to have a goal in mind _before_ entering the game (precisely because we’re not good enough at planning while we have to perform all the other cognitive tasks at the same time). Still, we definitely need to be able to adjust our plan once we encounter a deviation in the game state that makes our initial plan infeasible. People good at planning can come up with very efficient build orders or are able to optimally exploit a given map.

Conclusion
I have given you a short overview of some cognitive capabilities I consider very relevant in the context of StarCraft. On one hand, it is a very narrow selection and I haven’t even touched concepts like perception or learning yet. On the other hand, each of the mentioned cognitive capabilities is a research field in its own right and there is no way to give them justice in a blog post. If you are interested, please take this post as a pointer for your own investigations. I am sure you can also benefit as a human player.

If we want to create a successful bot, we need to master the presented cognitive capabilities in one way or another. In a next post, I will attempt to map the cognitive capabilities to attributes of an artificial intelligence, compare strengths and weaknesses and hopefully lay out a road map to cover current deficiencies of bots.

Disclaimer
Although I did deal with this subject in-depth for a good while, I am not a Neuroscientist. Apart from a 3-digit number of published scientific papers that I read I do also rely on Wikipedia as a source.



50 pts Copper League
Jett.Jack.Alvir
Profile Blog Joined August 2011
Canada2250 Posts
November 23 2016 15:54 GMT
#2
If you are trying to emulate human intelligence using those 5 processes, which would you say would be the most difficult for an artificial intelligence bot?

I can already assume attentional control and working memory are easy to design, unless you intentionally design your bot to get distracted by shiny objects and a memory of a goldfish. In fact, this might be where an artificial intelligence bot will be superior to human intelligence (aside from having automaton 1000 micro)

Reasoning will have limitations as the game goes longer, so that might be challenging for your bot. In super long games where all upgrades/units are unlocked, trying to use reasoning might become an exercise in futility.

Problem solving will be the greatest challenge for an AI, specifically ad-hoc problem solving. Your AI could have a plethora of problems it has encountered and solved, but can it account for every problem ever encountered in a game of SC? Because one of the truly remarkable capacities of humans is to 'think on our feet', which is a realm an AI might struggle with.
imp42
Profile Blog Joined November 2010
398 Posts
November 23 2016 16:20 GMT
#3
you mostly answered it yourself, here is my view as a preview of the next post:

Attentional Control: needs to be modeled for AI as well. Yes, a computer can simultaneously handle 3 incoming drops with ease and split the army perfectly, but at a lower level we have to deal with how this is done. Which process gets priority and access to CPU cycles?

Working Memory: not an issue for an AI.

Reasoning: Computers are good at reasoning if they are given a universe in which to reason. See e.g. theorem provers written in logical programming languages.

Problem solving requires the crucial abilities to generalize and to transfer.

Planning requires the crucial abilities to extrapolate and imagine.

So yes, the last two contain the real challenges and current Neural Nets are not the solution to those challenges IMO.

PS: you will probably laugh but I have spent several days thinking about how to make a computer discover the concept of addition (unsupervised), and I haven't come up with a good answer yet

E.g.: given the fact that 1 + 1 = 2, how can a computer extrapolate that 2 + 1 = 3 and how can it generalize that the concept holds for any natural numbers?
50 pts Copper League
Jett.Jack.Alvir
Profile Blog Joined August 2011
Canada2250 Posts
November 23 2016 16:50 GMT
#4
Yeah I assumed as much.

So for attentional control, it seems you are limited by hardware. If you had access to Google's data centres across the world, this shouldn't be a problem. Theoretically you can have your AI pay attention to everything.

It's funny what we take for granted as humans. Computers can do things no human could ever think of, but the simple things that we assume are easy are huge obstacles for computers.
imp42
Profile Blog Joined November 2010
398 Posts
November 23 2016 17:22 GMT
#5
On November 24 2016 01:50 Jett.Jack.Alvir wrote:
It's funny what we take for granted as humans. Computers can do things no human could ever think of, but the simple things that we assume are easy are huge obstacles for computers.


One thing I have observed, is that many people including well-known researchers make the "mistake" to judge the performance of a computer by a problem that is very biased towards humans.

For example: Neural Networks have been trained to play Frostbite better than humans. But some authors criticize that the Neural Networks need much more learning iterations (hundreds of thousands of iterations) than a human (a couple of minutes) to master the game.

What they forget when doing this comparison is the fact that any game produced by humans is specifically designed to suit the human brain. That is, it follows a set of rules that we're pretty much hardwired to know (e.g. human babies already have some concept of temporal causality).

In Pong, we just know that if a ball goes in a certain direction, it will follow that direction in the next frame, unless it hits an obstacle. In Frostbite we know that if birds kill you and fish give extra points, they will continue to do so in the next frame.

But what if we designed a game of Frostbite, where birds and fish reward or kill the player based on the pixel position of the floating ice multiplied by the game score or in-game frame count minus (randomly changing) color value of the polar bear?
Well, it wouldn't make a difference to the Neural Net but the human brain would struggle to the point it would under-perform even after hundreds of thousands of iterations.

Yes, we humans have a pretty good general problem solving tool in our heads. But at the same time it is also highly specialized to deal with problems that have been historically relevant to survival in our physical world and it will often fail miserably outside that domain.
50 pts Copper League
angrybacon
Profile Joined March 2010
United States98 Posts
November 24 2016 05:34 GMT
#6
On November 24 2016 02:22 imp42 wrote:
But at the same time it is also highly specialized to deal with problems that have been historically relevant to survival in our physical world and it will often fail miserably outside that domain.


The same can be said of neural nets.
imp42
Profile Blog Joined November 2010
398 Posts
November 24 2016 06:28 GMT
#7
On November 24 2016 14:34 angrybacon wrote:
Show nested quote +
On November 24 2016 02:22 imp42 wrote:
But at the same time it is also highly specialized to deal with problems that have been historically relevant to survival in our physical world and it will often fail miserably outside that domain.


The same can be said of neural nets.


yes, maybe it can.
Still, comparing neural nets and human brains on the grounds of video games developed for us by us seems to me as valuable as comparing movement speeds of the fish and the bird when doing all the tests in the water.
50 pts Copper League
Please log in or register to reply.
Live Events Refresh
Maestros of the Game
17:00
Group Stage - Group B
Zoun vs ByuNLIVE!
ComeBackTV 1076
SteadfastSC669
IndyStarCraft 227
BRAT_OK 126
EnkiAlexander 76
Rex44
LiquipediaDiscussion
[ Submit Event ]
Live Streams
Refresh
StarCraft 2
SteadfastSC 669
IndyStarCraft 222
BRAT_OK 122
ProTech102
JuggernautJason84
Rex 42
CosmosSc2 41
StarCraft: Brood War
Britney 18467
Dewaltoss 147
sSak 23
NaDa 10
Counter-Strike
Stewie2K494
Heroes of the Storm
Liquid`Hasu422
Khaldor119
Other Games
Grubby2295
summit1g2263
tarik_tv1859
fl0m1151
Nathanias560
JimRising 327
C9.Mang0145
Livibee60
ViBE29
PPMD27
Maynarde5
Organizations
StarCraft 2
angryscii 12
Blizzard YouTube
StarCraft: Brood War
BSLTrovo
sctven
[ Show 22 non-featured ]
StarCraft 2
• musti20045 31
• StrangeGG 17
• davetesta10
• Reevou 2
• sooper7s
• AfreecaTV YouTube
• intothetv
• Kozan
• IndyKCrew
• LaughNgamezSOOP
• Migwel
StarCraft: Brood War
• HerbMon 35
• FirePhoenix16
• Pr0nogo 5
• STPLYoutube
• ZZZeroYoutube
• BSLYoutube
Dota 2
• masondota22184
League of Legends
• Doublelift5212
Counter-Strike
• Shiphtur182
Other Games
• imaqtpie688
• Scarra372
Upcoming Events
CranKy Ducklings
12h 22m
RSL Revival
12h 22m
GuMiho vs Cham
ByuN vs TriGGeR
Cosmonarchy
16h 22m
TriGGeR vs YoungYakov
YoungYakov vs HonMonO
HonMonO vs TriGGeR
Maestros of the Game
19h 22m
Solar vs Bunny
Clem vs Rogue
[BSL 2025] Weekly
20h 22m
OSC
1d
RSL Revival
1d 12h
Cure vs Bunny
Creator vs Zoun
Maestros of the Game
1d 19h
Maru vs Lambo
herO vs ShoWTimE
BSL Team Wars
1d 21h
Team Hawk vs Team Sziky
Sparkling Tuna Cup
2 days
[ Show More ]
Monday Night Weeklies
2 days
The PondCast
5 days
Online Event
6 days
BSL Team Wars
6 days
Team Bonyth vs Team Dewalt
BSL Team Wars
6 days
Liquipedia Results

Completed

Proleague 2025-09-02
SEL Season 2 Championship
HCC Europe

Ongoing

Copa Latinoamericana 4
BSL 20 Team Wars
KCM Race Survival 2025 Season 3
BSL 21 Qualifiers
ASL Season 20
CSL 2025 AUTUMN (S18)
LASL Season 20
RSL Revival: Season 2
Maestros of the Game
BLAST Open Fall 2025
BLAST Open Fall Qual
Esports World Cup 2025
BLAST Bounty Fall 2025
BLAST Bounty Fall Qual
IEM Cologne 2025
FISSURE Playground #1

Upcoming

2025 Chongqing Offline CUP
BSL Polish World Championship 2025: Warsaw LAN
BSL Season 21
BSL 21 Team A
Chzzk MurlocKing SC1 vs SC2 Cup #2
EC S1
BLAST Rivals Fall 2025
IEM Chengdu 2025
PGL Masters Bucharest 2025
Thunderpick World Champ.
MESA Nomadic Masters Fall
CS Asia Championships 2025
ESL Pro League S22
StarSeries Fall 2025
FISSURE Playground #2
TLPD

1. ByuN
2. TY
3. Dark
4. Solar
5. Stats
6. Nerchio
7. sOs
8. soO
9. INnoVation
10. Elazer
1. Rain
2. Flash
3. EffOrt
4. Last
5. Bisu
6. Soulkey
7. Mini
8. Sharp
Sidebar Settings...

Advertising | Privacy Policy | Terms Of Use | Contact Us

Original banner artwork: Jim Warren
The contents of this webpage are copyright © 2025 TLnet. All Rights Reserved.