Using AI to Help with Intel Analysis

[Kestrel25]

I'm playing a PBEM game where it's about July, 1942. There are about 1,300 intel, ops, and combat reports. I imported the text files to Notion AI, separated by report type. I then give Notion AI commands like "identify all locations by date where heavy volume of radio transmissions appear" or "identify all dates where CV Kaga is named".

I'm working on building a table (or tables) for order of battle analysis and perhaps even for submarine or pilot performance (not sure if sufficient data exists here).

What are thoughts about using these tools? And what other ways can AI be used to help with the large volume of information?

[CaptBeefheart]

Have you tried Tracker? That may cover what you're looking for, although it would be interesting if the AI gives you the info in a more useful format.

Cheers,
CB

[Kestrel25]

I have tracker, which is great to put all the reporting in an organized, readable format. But it's not that great to make analytic judgments, or even to track OOB for a specific location over time.

[Chris21wen]

There is a small app called Intel Monkey somebody wrote which I think will meet your needs but I've never used it as I don't think such intel detail is needed. Other plays use it all the time so it must be usefull to them.

[Trugrit]

I run Tracker and Combat Reporter at the same time side by side.
That is all I’ve ever needed.

They both have Intel info categorized.

They are good programs…..You can always give them a test run to see if
they work for you before you write your own program.

Combat Reporter:
https://www.matrixgames.com/forums/view ... 3#p2315353
https://sites.google.com/site/witpaecombatreporter/
If you can get it to run….it is tricky to get running.

Tracker:
https://www.matrixgames.com/forums/view ... 3&t=165129
https://sites.google.com/site/witptracker
Also tricky to get running.

tracker23.jpg (227.47 KiB) Viewed 771 times

[Kestrel25]

Thanks. I got them mixed up - I have Combat Reporter but have been unable to get tracker working so I don't know what functions I'm missing there. Here's an example of the feedback I received when I asked the AI to summarize reporting about a specific location (I redacted the identifying information in case my opponent snoops here ):

On April 29, 1942, [redacted] was located at [redacted], indicating a military presence in the area. Additionally, [redacted] was also reported to be located at [redacted] on the same date. However, aside from these reports, there is no further information or details provided about [redacted] specifically.

Although there are mentions of radio transmissions being detected at [redacted] on April 19, 1942, and June 1, 1942, there is no information provided regarding the content or significance of these transmissions. Similarly, on May 1, 1942, there is a report of radio transmissions being detected at [redacted], but no further information is given about the nature of these transmissions.

In summary, while [redacted] is mentioned in several reports throughout the given text, there is no detailed information or activity mentioned about [redacted] itself. If you have any further questions or concerns, please let me know.

- [redacted] was located at [redacted] on April 29, 1942
- [redacted] was reported to be located at [redacted] on April 29, 1942
- Radio transmissions were detected at [redacted] on April 19, 1942, June 1, 1942, and May 1, 1942, but no information was provided about the content or significance of these transmissions

I can also ask the AI for general information like "summarize Japanese activity in the Malaysia region" and it'll write a report based on unit locations. It's not perfect but does give me a better idea of what the reporting is telling me.

[BBfanboy]

Is [redacted] reporting a problem with their fresh water condensers?

[rook749]

There is a small app called Intel Monkey somebody wrote which I think will meet your needs but I've never used it as I don't think such intel detail is needed. Other plays use it all the time so it must be usefull to them.

I use Intel Monkey with Tracker and Combat Report when I do PBEM games. Its a great tool but I don't use it when I play the AI but against a Human it really allows you to use the Allied Intel to full effect.

[Kestrel25]

Quick update: I managed to get IntelMonkey to work. Now I can reduce the amount of text the AI has to review by filtering based on IM's outputs. I can then copy those tables into Notion AI and allow it to assess patterns.

[BBfanboy]

Does your AI know that the strategy is to be set within game parameters, not real life ones? In the game, scoring VPs is important. And the Allied timeline goal is to be within bombing range of Japan in 1944 to harvest PP from bombing industries, while the Japanese goal is to prevent this while scoring PP from destroying ships, aircraft and LCUs.

[Kestrel25]

Not inherently. But the AI can informed about it and take it into account.

[MarkShot]

I have been working with these GPTx engines.

Their recall of information is truly amazing especially on an old game such as this. This is because their many training corpus has already been produced, and cutoff in time when they were produced like 2021. Few crawl the Web like search engine spiders. Training hardware and execution hardware are distinct; unlike your desktop.

So, if you want information and insights about the game, they can be useful. But remember, no matter what they say, they never played it. These are the insights of expert human players who mainly posted here. There are open source curated corpus from both Wikipedia and Reddit. I am not sure Matrix made the general Web cut off.

Now, the third party tools we use while playing turn data (atoms, if you will) into information (polymer chains, something useful). Why do we want information? Because from information, we can reason, draw conclusions, and do something novel. GPT3 has non of those capabilities. It can write a poem or paint as it follows patterns and templates, it has built. But it cannot synthesize a new art form poetry/painting fusion.

So, can it help you play this game? Only helping you find information quicker than a search. Deep learning neural nets could play this game better than any human, and games are perfect since the net needs some a form of scoring or validation feedback and of, course, games provide this. It took 10,000 GPUs and 300 days to produce GPT3 model. Given that those who are doing the final release are unpaid. I doubt anyone will invest $50M USD to train the ultimate ML player of WITP-AE. So, you are on your own with strategy, and PBEM is your best option for a challenge despite GPT4 or GPT5.

---

Another thread asks if GPT3 makes this forum obsolete? Yes and no. It (LLM NLP) is the best search, you can find. But it does not play, it is leeching off of human players. We have the content producer problem. When those who post are discouraged, because of the speed and volume AI can spit out, raw information will dry up. And without those players, you are going to find GPT6 to be quite dumb about things like WITP-AE. This is only one problem this technology poses is the death of human creativity, not necessarily supplanted by machine creativity; just real fast plagiarism.