CRAZY idea
Moderator: Shannon V. OKeets
CRAZY idea
I was thinking about AI for this game... and thought why bother? instead of embedding AI in the game why not have AI play the game just like a human would? sounds interesting so I asked AI if it's possible and this is what I got
Copilot said:
Hypothetically, yes.
If a future version of Copilot had the ability to:
Launch and interact with Windows applications,
Read the World in Flames game state from the screen or memory,
Click buttons, move units, and make decisions,
then there is no fundamental reason it couldn't act as an AI opponent.
In fact, World in Flames may be a particularly good fit for an LLM-based opponent because much of the challenge is strategic planning, resource allocation, diplomacy, production, and long-term operational decisions rather than fast reflexes.
That said, there are some obstacles:
Rules complexity: World in Flames is one of the most complex grand-strategy wargames ever published.
Long-term planning: A competent opponent must plan years of game time ahead.
Search depth: Strong play often requires evaluating many possible moves and counter-moves, which may require traditional game-AI techniques in addition to an LLM.
State tracking: The AI would need an accurate representation of thousands of units, production choices, and map positions.
A likely architecture would be:
The game engine provides the legal moves.
A planning/search engine evaluates alternatives.
An LLM serves as the strategic commander, setting priorities and explaining its decisions.
So the answer is not "Copilot is an AI opponent today," but rather "if Copilot could directly control the game, it could probably serve as an opponent—and potentially a pretty interesting one—provided it had access to the complete game state and enough reasoning resources."
Personally, I suspect an AI designed specifically for World in Flames could eventually become a very strong player because the game rewards strategic reasoning more than mechanical execution.
Copilot said:
Hypothetically, yes.
If a future version of Copilot had the ability to:
Launch and interact with Windows applications,
Read the World in Flames game state from the screen or memory,
Click buttons, move units, and make decisions,
then there is no fundamental reason it couldn't act as an AI opponent.
In fact, World in Flames may be a particularly good fit for an LLM-based opponent because much of the challenge is strategic planning, resource allocation, diplomacy, production, and long-term operational decisions rather than fast reflexes.
That said, there are some obstacles:
Rules complexity: World in Flames is one of the most complex grand-strategy wargames ever published.
Long-term planning: A competent opponent must plan years of game time ahead.
Search depth: Strong play often requires evaluating many possible moves and counter-moves, which may require traditional game-AI techniques in addition to an LLM.
State tracking: The AI would need an accurate representation of thousands of units, production choices, and map positions.
A likely architecture would be:
The game engine provides the legal moves.
A planning/search engine evaluates alternatives.
An LLM serves as the strategic commander, setting priorities and explaining its decisions.
So the answer is not "Copilot is an AI opponent today," but rather "if Copilot could directly control the game, it could probably serve as an opponent—and potentially a pretty interesting one—provided it had access to the complete game state and enough reasoning resources."
Personally, I suspect an AI designed specifically for World in Flames could eventually become a very strong player because the game rewards strategic reasoning more than mechanical execution.
Re: CRAZY idea
wow Copilot când do to it but Cluade seems capable of it "evelopers can direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking buttons, and typing text. "
Re: CRAZY idea
"Claude 4.5 Sonnet just hit a 61% success rate on computer use benchmarks, up from 42% with the previous model. That's not an A+, but hey—it's passing!
Computer use means AI can interact with software through the same interface you use. It sees what's on your screen through screenshots. It figures out where to click based on what it sees and what you asked it to do. It moves the cursor, clicks, types, waits for the screen to update, takes another screenshot, and continues until it completes the task."
Computer use means AI can interact with software through the same interface you use. It sees what's on your screen through screenshots. It figures out where to click based on what it sees and what you asked it to do. It moves the cursor, clicks, types, waits for the screen to update, takes another screenshot, and continues until it completes the task."
Re: CRAZY idea
Nothing is crazy at this point! Matrix needs to do something. Waiting for the mythical Steve to get anything done is not working--I think there's a better chance of Bigfoot being real.
Re: CRAZY idea
An AI program needs files to draw it's information from. That isn't only the rules, maps and units, but also a lot of gameplay. We all know that certain attacks you can affort to take the risk and other you don't with the same odds. Now, if it would be possible for the program to send those games on the internet, the AI programs might pick that up and learn from it. The idea isn't crazy at all.
Peter
- juntoalmar
- Posts: 703
- Joined: Sun Sep 29, 2013 2:08 pm
- Location: Valencia
- Contact:
Re: CRAZY idea
sorry, I think nowadays AI is not powerful enough to work with the convoy system user interface 
(my humble blog about wargames, in spanish) http://cabezadepuente.blogspot.com.es/
Re: CRAZY idea
What about if the AI "digested" & operated on the game file? No need to worry about reading the screen, mouse clips, keyboard clicks.
For this to work the AI would have to also "digest" MWIF input data files, understand & apply the rules correctly.
And, significantly understand the format of the MWIF game file.
There would have to be some ability to interact with (talk with) the AI to resolve cases when they made illegal moves, which at first could be a significant number.
I could even see cases where the AI argued that the move they made was legal and your understanding of the rules was wrong.
To do this, I would go back to the old riddle of how does one eat an elephant? The answer is one bite at a time.
Meaning, maybe break this approach down into smaller steps (e.g., attack planning, convoy planning, land moves, land attacks). Steps that initially would be fed by decisions made by humans but with those decision replaced with AI decision making once the fundamental AI steps were working reasonably well.
For this to work the AI would have to also "digest" MWIF input data files, understand & apply the rules correctly.
And, significantly understand the format of the MWIF game file.
There would have to be some ability to interact with (talk with) the AI to resolve cases when they made illegal moves, which at first could be a significant number.
I could even see cases where the AI argued that the move they made was legal and your understanding of the rules was wrong.
To do this, I would go back to the old riddle of how does one eat an elephant? The answer is one bite at a time.
Meaning, maybe break this approach down into smaller steps (e.g., attack planning, convoy planning, land moves, land attacks). Steps that initially would be fed by decisions made by humans but with those decision replaced with AI decision making once the fundamental AI steps were working reasonably well.
Ronnie
- juntoalmar
- Posts: 703
- Joined: Sun Sep 29, 2013 2:08 pm
- Location: Valencia
- Contact:
Re: CRAZY idea
yes, that's reasonable. The map with the counters are, after all, a visual representation of the game file for humans.
(my humble blog about wargames, in spanish) http://cabezadepuente.blogspot.com.es/
Re: CRAZY idea
(1) Below is a reposting of my most recent entry to my on-going AAR A Competitive GW (Solo) Scenario by the Numbers. on page 105 (don't know how to reference post # on that page).
(2) While I do (have) dabbled in human-to-human play, the majority of my play has been competitive solo; if such a thing exist.
(3) Both in addition & support of my competitive solo play I've developed a number tools that assist me if my tactical and even strategic decisions.
(4) These tools would also be of benefit in human to human play.
(5) For example, I can now easily summarize land, sea & air moves that occurs during an impulse.
(6) So in competitive human-to-human play I could use such to see exactly what moves my opponent made so I wouldn't be surprised by any unintentional or even intentional "sneaky" moves.
(7) In a friendly human-to-human game this could also serve as a summary to my opponent of my moves.
(8) My toolset is built primarily on EXCEL but also includes a python application that quickly processes the game file to generate a "on-map" and "full" counters data (csv) files.
(9) I tried to accomplish that in EXCEL using VBA (Visual Basic for Applications) ; but run times were 900 seconds compared to a fraction of a second in python & the VBA would occasionally lock up my computer; so I gave up on that.
(10) One upside is that none of my EXCEL tools (or workbooks) include any VBA modules; i.e., all of them use EXCEL formulas.
(11) This means all my EXCEL workbooks have .xlsx (vs .xlsm) extensions.
(12) .xlsm extension are for workbooks including VBA & present their own problems concerning computer and internet anti- virus/malicious code scanners.
(13) The BOA Sub Matrix below was automatically generated from on-map csv counters file extracted by MAGIC (python executable I built) then entered into MAGIC_EXCEL.xlsx (a multi-sheet EXCEL workbook that I also built).
(14) This workbook includes a number of sheets for reports (land, naval, air) that I've included in the reference AAR (above), all of which were automatically generated (well, mostly automatically generated in that I had to filter on which units I wanted in the various reports).
(15) I used the BOA Sub Matrix; specifically, G/L (Gain/Loss) values to assist me in determining which action Germany should take (i.e., combine or land) and whether or not Italy such search with her Subs in the North Atlantic & Cape Verde Basin.
(16) So when I was talking in my previous post about eating an elephant one bite at time, I'm trying to use this post as an example.
(17) For example; if one could feed my current game file & any needed supporting data files into AI (however, one would do that); then would it possible for the AI to go through the same extractions & calculations that I just did with Python and EXCEL; but far faster, to arrive at the same (or different) decisions that I did concerning German action type (i.e., combine vs land) and whether or not Italy should search in the North Atlantic and/or Cape Verde Basin?
(18) Also, I have a Land Combat attack planning workbook (reports from not shown here; but extensively presented in my AAR), that I use to assist me in making my land move & land combat decisions.
(19) Again, it would be nice if AI feed the appropriate files and player learning experience could make such decisions or at least make recommendations.
(20) I'll go back to CanInf's original post & title, is all of this a "CRAZY idea" ???
(21) One can dream can't they?
Turn 19. Sep/Oct 1942. Axis #13. Atlantic Theater.
Combat Logs. BOA Sub Matrix. North Atlantic. Cape Verde Baisn.
(2) While I do (have) dabbled in human-to-human play, the majority of my play has been competitive solo; if such a thing exist.
(3) Both in addition & support of my competitive solo play I've developed a number tools that assist me if my tactical and even strategic decisions.
(4) These tools would also be of benefit in human to human play.
(5) For example, I can now easily summarize land, sea & air moves that occurs during an impulse.
(6) So in competitive human-to-human play I could use such to see exactly what moves my opponent made so I wouldn't be surprised by any unintentional or even intentional "sneaky" moves.
(7) In a friendly human-to-human game this could also serve as a summary to my opponent of my moves.
(8) My toolset is built primarily on EXCEL but also includes a python application that quickly processes the game file to generate a "on-map" and "full" counters data (csv) files.
(9) I tried to accomplish that in EXCEL using VBA (Visual Basic for Applications) ; but run times were 900 seconds compared to a fraction of a second in python & the VBA would occasionally lock up my computer; so I gave up on that.
(10) One upside is that none of my EXCEL tools (or workbooks) include any VBA modules; i.e., all of them use EXCEL formulas.
(11) This means all my EXCEL workbooks have .xlsx (vs .xlsm) extensions.
(12) .xlsm extension are for workbooks including VBA & present their own problems concerning computer and internet anti- virus/malicious code scanners.
(13) The BOA Sub Matrix below was automatically generated from on-map csv counters file extracted by MAGIC (python executable I built) then entered into MAGIC_EXCEL.xlsx (a multi-sheet EXCEL workbook that I also built).
(14) This workbook includes a number of sheets for reports (land, naval, air) that I've included in the reference AAR (above), all of which were automatically generated (well, mostly automatically generated in that I had to filter on which units I wanted in the various reports).
(15) I used the BOA Sub Matrix; specifically, G/L (Gain/Loss) values to assist me in determining which action Germany should take (i.e., combine or land) and whether or not Italy such search with her Subs in the North Atlantic & Cape Verde Basin.
(16) So when I was talking in my previous post about eating an elephant one bite at time, I'm trying to use this post as an example.
(17) For example; if one could feed my current game file & any needed supporting data files into AI (however, one would do that); then would it possible for the AI to go through the same extractions & calculations that I just did with Python and EXCEL; but far faster, to arrive at the same (or different) decisions that I did concerning German action type (i.e., combine vs land) and whether or not Italy should search in the North Atlantic and/or Cape Verde Basin?
(18) Also, I have a Land Combat attack planning workbook (reports from not shown here; but extensively presented in my AAR), that I use to assist me in making my land move & land combat decisions.
(19) Again, it would be nice if AI feed the appropriate files and player learning experience could make such decisions or at least make recommendations.
(20) I'll go back to CanInf's original post & title, is all of this a "CRAZY idea" ???
(21) One can dream can't they?
Turn 19. Sep/Oct 1942. Axis #13. Atlantic Theater.
Combat Logs. BOA Sub Matrix. North Atlantic. Cape Verde Baisn.
Ronnie
Re: CRAZY idea
(1) I also wanted to show off, well brag on, some (more) examples of MWIF play support functionality that I've built & iterated on over the past 10+ years.
(2) Magic Excel: Atlantic/Pacific calculations are fully automated based on extracted unit game states (exacted by Python MAGIC).
(3) With two exceptions, which are that I have to manually enter the weather (i.e., Fine, Rain, Snow, Blizzard, Storm) in each of the 6 weather zones.
(4) The spreadsheet then calculates the air-to-sea weather multiplication factor (Wx) for each sea area based on weather zone.
(5) The second manual set of inputs is A2S(R), which are the air-to-sea factors of air units in position and likely to react to sea area [0-box] is naval search is initiated.
(6) All other values are automatically calculated.
(7) BOA Calculator: This is a work in progress, porting some functionality from MAGIC (python) to MAGIC (excel) now that MAGIC (excel) is operating directly on unit game states.
(8) The block of cells in the lower right within the red box are data that I will copy and paste directly into the cells of my AAR combat log workbook.
(9) Everything is automatically calculated except for the location code of subs (or ships) in port, gain & loss values.
(10) In the upper right, the user enters the UID (unit IDs) of the sub Gps and/or warships that are applicable threats to allied CPs.
(11) The rest for the sub/ship is calculated with, as stated previously, the location code for subs/ships in port (I'm working on that).
(12) Location code for ships at sea are [Sea_Area_ID.Sea_Box] and for subs/ships in port A=Kiel, B=La Spezia, C=Western France along the Bay of Biscay, D=Along the Baltic.
(13) Of course, this needs to be more generalized and is something I plan to address when it gets the priority.
(14) MAGIC DUST: Magic Dust contains import naval, land, air and other MWIF input data files. It also includes "hc" tab which I'm generating for city/port/sea area locations, abbreviations and default theaters/command (e.g., Western Front, Eastern Front, Pacific).
(15) I currently used 12 theaters & 2 commands (Axis, Allied).
(16) Attack Planning, 2D10 Calculator: The 2D10 CRT calculator (bottom half of image) is the earliest tool that I built and is all player input driven; well except for the resultant calculations.
(17) It includes two other tabs that I use for analysis and for transferring to my Land Combat Attack Planning workbook.
(18) Well, it's kind of a chicken and egg with these two.
(19) In the 2D10 AP workbook I first enter the relevant attacker and defender inputs (everything except what's highlighted in orange, which is calculated).
(20) I then input the calculated odds, which other player inputs, to my 2D10 CRT calculator to get PWIN and other probabilities.
(21) I then go to the report tab, copy and paste those stats into the (blue) right section of my 2D10 AP worksheet.
(22) The unhighlighted (ugly) sector of to the far right (column AL at later), I then copy and paste directly into my AAR EXCEL worksheet (L, short for log).
Magic Excel - Atlantic/Pacific Tab. BOA Calculator Tab. Magic Dust Land Combat Attack Planning, 2D10 Calculator
(2) Magic Excel: Atlantic/Pacific calculations are fully automated based on extracted unit game states (exacted by Python MAGIC).
(3) With two exceptions, which are that I have to manually enter the weather (i.e., Fine, Rain, Snow, Blizzard, Storm) in each of the 6 weather zones.
(4) The spreadsheet then calculates the air-to-sea weather multiplication factor (Wx) for each sea area based on weather zone.
(5) The second manual set of inputs is A2S(R), which are the air-to-sea factors of air units in position and likely to react to sea area [0-box] is naval search is initiated.
(6) All other values are automatically calculated.
(7) BOA Calculator: This is a work in progress, porting some functionality from MAGIC (python) to MAGIC (excel) now that MAGIC (excel) is operating directly on unit game states.
(8) The block of cells in the lower right within the red box are data that I will copy and paste directly into the cells of my AAR combat log workbook.
(9) Everything is automatically calculated except for the location code of subs (or ships) in port, gain & loss values.
(10) In the upper right, the user enters the UID (unit IDs) of the sub Gps and/or warships that are applicable threats to allied CPs.
(11) The rest for the sub/ship is calculated with, as stated previously, the location code for subs/ships in port (I'm working on that).
(12) Location code for ships at sea are [Sea_Area_ID.Sea_Box] and for subs/ships in port A=Kiel, B=La Spezia, C=Western France along the Bay of Biscay, D=Along the Baltic.
(13) Of course, this needs to be more generalized and is something I plan to address when it gets the priority.
(14) MAGIC DUST: Magic Dust contains import naval, land, air and other MWIF input data files. It also includes "hc" tab which I'm generating for city/port/sea area locations, abbreviations and default theaters/command (e.g., Western Front, Eastern Front, Pacific).
(15) I currently used 12 theaters & 2 commands (Axis, Allied).
(16) Attack Planning, 2D10 Calculator: The 2D10 CRT calculator (bottom half of image) is the earliest tool that I built and is all player input driven; well except for the resultant calculations.
(17) It includes two other tabs that I use for analysis and for transferring to my Land Combat Attack Planning workbook.
(18) Well, it's kind of a chicken and egg with these two.
(19) In the 2D10 AP workbook I first enter the relevant attacker and defender inputs (everything except what's highlighted in orange, which is calculated).
(20) I then input the calculated odds, which other player inputs, to my 2D10 CRT calculator to get PWIN and other probabilities.
(21) I then go to the report tab, copy and paste those stats into the (blue) right section of my 2D10 AP worksheet.
(22) The unhighlighted (ugly) sector of to the far right (column AL at later), I then copy and paste directly into my AAR EXCEL worksheet (L, short for log).
Magic Excel - Atlantic/Pacific Tab. BOA Calculator Tab. Magic Dust Land Combat Attack Planning, 2D10 Calculator
Ronnie
Re: CRAZY idea
(1) I promise this will be the last of my examples, unless ...
(2) Magic Python: Used to generate to extract unit states from the game file and save them to csv file, which is then imported into EXCEL.
(3) Magic EXCEL: Unit game states are extracted by Magic (python) at two points (0=e.g., beginning of impulse) & (1=e.g., after land moves or end of impulse), copied and pasted into Magic Excel.
(4) Columns O - AR are then automatically calculated using the equations shown in the control/equations tab.
(5) AAR EXCEL: This is where I log and keep track of my game (in the "L" tab).
(6) I currently have 57 different (variable) log types on which I capture detailed data.
(7) The "LGR" tab is where I produce a number of combat logs report that I shown in my AAR.
Magic (Python) Magic Excel AAR Excel
(2) Magic Python: Used to generate to extract unit states from the game file and save them to csv file, which is then imported into EXCEL.
(3) Magic EXCEL: Unit game states are extracted by Magic (python) at two points (0=e.g., beginning of impulse) & (1=e.g., after land moves or end of impulse), copied and pasted into Magic Excel.
(4) Columns O - AR are then automatically calculated using the equations shown in the control/equations tab.
(5) AAR EXCEL: This is where I log and keep track of my game (in the "L" tab).
(6) I currently have 57 different (variable) log types on which I capture detailed data.
(7) The "LGR" tab is where I produce a number of combat logs report that I shown in my AAR.
Magic (Python) Magic Excel AAR Excel
Ronnie


