Malo sam sinoc razmisljao o AI i kuda sve to vodi... I dosao sam do sljedeceg zakljucka...
Dakle, LLM mi i dalje nema puno smisla osim za komunikaciju covjeka i kompjutera, za to je dobar, no u pozadini on se mora povezivati sa specijaliziranim AI modelima za druge stvari... Ne, ne pricam o Dall-E i generiranju slika i slicnim pizd*arijama... Pricam o realnoj primjeni gdje AI moze stvarno pokazati svu svoju raskos...
Ono sto predvidjam u nekoj relativno bliskoj buducnosti tipa sljedecih 10-15 godina su raznorazne letjelice... Dizanje masovnog prometa u trecu dimenziju... Znaci nikakvi robotaxiji i pizdarije, pricam o dron-taxijima kakve radi
EHang... Nadalje, osim robotaxija i opcenito malih osobnih letjelica, bit ce tu i hrpetina dostavnih dronova... Pricamo o valjda preko 100 milijuna malih dronova svih velicina i oblika koji ce sluziti za dostave svega i svacega...
Realno kada sada pogledamo nebo u USA, tamo leti mozda par tisuca aviona i helikoptera svih velicina... I postoji kontrola leta koja sve to nadgleda da ne bude problema... To su onda medjugradski i medjudrzavni letovi... No, ovo sto ja spominjem gore su gradski letovi, dakle, trebat ce organizirati kontrolu leta malih gradskih letjelica... I to milijune njih, a to ljudi vise nece moci kontrolirati...
I tocno tu vidim realnu primjenu AI sustava... I to ne LLM laprdala, jer oni ne mogu to raditi, vec uvezanih AI sustava, a tu onda pricamo stvarno o svemu i svacemu... Poanta je da imamo realtime decision-making, predikcije trajektorija (na bazi milijuna letjelica, svaka sa svojih brdo parametara), crash collision avoidance, itd...
Upisao sam u Google sta bi mi za to trebalo i dobio sljedece... Cisto da malo vidite da AI nije samo LLM...
1. Conflict Detection & Resolution (CD&R)
- Deep Reinforcement Learning (DRL): Considered highly suitable for decision-making, DRL agents can learn to maximize safety by maintaining separation, issuing speed, altitude, or heading changes.
- Supervised Learning (Random Forest/SVM): These can be used to classify potential conflicts and predict optimal, predefined resolution strategies based on historical controller data.
- Deep Q-Learning: A model-free RL technique ideal for learning complex, sequential decision-making policies from historical simulation data.
2. Trajectory Prediction & Flow Management
- Graph Neural Networks (GNNs): Specifically useful for modeling air traffic networks, GNNs can capture complex spatial relationships between aircraft, sectors, and airports.
- Recurrent Neural Networks (RNNs) / LSTM: Effective for handling sequential, time-series data to predict 4D aircraft trajectories.
- Gradient Boosting Decision Trees (GBDT): Effective for predicting Estimated Time of Arrival (ETA) and managing flow to avoid sector congestion.
3. Surface Movement & Monitoring
- Convolutional Neural Networks (CNNs): Used in Computer Vision to monitor airport surface movement via cameras, track aircraft, and detect potential runway incursions.
- DBSCAN (Density-Based Spatial Clustering): Used for identifying hotspots and analyzing traffic flow patterns without needing labels.
4. System Architecture
- Hybrid Rule-Based/AI Systems: Combine machine learning with traditional, rigid, rule-based systems to ensure a "safety net" (e.g., ensuring collision avoidance regulations are never violated by the AI).
- Explainable AI (XAI) / Knowledge Graphs: Used in projects like SESAR TAPAS to ensure that AI decisions can be understood and trusted by human controllers.
BTW., da, izgleda da mi je Gemini koji je integriran u Google Search izlaprdao napokon nesto korisno...
Al realno, to mu je i namjena, sad je samo pitanje sta je od svega ovoga tocno...