Matiz the Math Wiz
A slighter deep dive into Matiz solution, setup and considerations
kulbir minhas
2/22/20244 min read
For first time readers of this topic, Please refer to “AI in Poker” for background of Matiz.
Matiz is a deep Adversarial collaborative neural network, which is trained to optimize certain activities, within the defined workflow and constraint. In the following example it is trained to play poker.
How Does Matiz Work?
Let's break down how Matiz does its magic.
First off, instead of trying to plan out the entire poker game beforehand, Matiz takes a more dynamic approach. It treats Poker like a series of changing situations, and it's super clever at adapting its strategy on the spot.
The secret sauce is in its "re-solving" trick. At each decision point, Matiz looks at what's happening in the game right then and there. This means it's not stuck to a fixed plan and it's making smart moves based on the current state of play.
To make decisions quickly, Matiz doesn't overthink things. It limits how far into the future it looks and uses a neural network to quickly figure out how valuable the game is at that moment. So, it's like its playing poker with a combination of street smarts and lightning-fast calculations.
Now, here's where it gets even cooler. Matiz learns from experience, thanks to deep learning. It's been trained on millions of poker scenarios, so it knows the game inside out. It's not just a rule follower, it has got some serious game sense.
And it's not just a poker genius. In real-world tests against the pros, Matiz can handle the chaos of real-world situations where information is incomplete, making it a versatile problem-solver.
Why Poker?
Poker is a game which the author loves and wanted to simulate the scenario and understand if AI can outperform humans.
According to NewScientist; “Poker is an excellent game for testing artificial intelligence,” says Michael Bowling, head of the Computer Poker Research Group at the University of Alberta, Canada.
It is a game of probability, luck, skill and the amount of computation skill AI Holds, I believe it gives it the edge.
There are elements like experience, exposure and creativity (bluff, variation in calls) etc. added with human endurance makes it an exciting match.
The following are some of the factors (parameter domains), which Matiz considers:
User Classification
Betting Pattern Classification
Base
Trend (Tilt etc.)
Position Evaluation
Card evaluation
Range Evaluation
Betting/probability Evaluation
Facial Tells and Other Biometrics
This experiment was set up with the following primary assumptions, it is possible to vary it, but this solution gives an ideal ground for AI to learn, detail results and comparative study will be published in subsequent papers
Primary Setup/Constraint:
No Limit Hold’em
No Ante, Standard SB/BB and Dealer Setup
Initial Starting Money (40-time BB)
No Rebuys
6 players' table
Boundary/Setups/Governance:
The models are trained to learn multistage (flop, turn and river)
The models are not explicitly trained order of cards
But any show of cards, adjudicator (network) tells who wins
Model over time learns the order of card
The models are collaborative in nature, they share the information
To speed up the learning process
Post processing model is ready to play with humans
Post model training, it was further exposed the following iteration and learning from each iteration included in the next iteration.
Iterations
Standard setup – AI Only
Standard setup – AI + human (AI and Human are aware they are playing against each other)
Machine/Human identity is Known
Standard setup – AI + human (Machine/Human identity is Undisclosed)
AI + Human (facial data)
AI + Human (facial data + real time biometrics)
Notes
The last 3 scenarios are where human + AI don’t know who they are playing against.
Humans + AI Avatars are playing against
AI Avatars are learning to simulate
Facial expression
NN model is trained to review the expression and simulate it
AI is also trained to simulate the biometrics (autogenerated) to mimic human behaviour
Scenario 4 and 5 are still ongoing test trials
Potential Next Steps
Poker:
Model Fine Tuning
AI Duel – Competing with other AI Poker solution
AI + Best in Poker Duel – Bring in best in Poker + AI together
Potentially WSOP Grand Slam (to be explored)
General:
The primary objective of Matiz is able to make the decisions, better than human at much improved speed.
This solution is not going to be specific to Poker
What Matiz is aiming for is super AI platform, able to solve any statistical problem, given enough information is available and make it more efficient
Users will be able to use this platform and just plug and play to solve their specific personal/business needed
Outcome:
Matiz is still in advanced stage of testing and model validation and fine tuning. Results are currently in process of internal review and will be shared for wider technical review.
It's designed to be a tough cookie. Matiz is desinged to learn and adapt, so any gaps/weakness is model, it is designed to adapt and improve, Whether it's at the poker table or facing real-world challenges, it's built to stand its ground. Its ability to learn and mimic Human creativity make it real poker master/trainer.
So, in a nutshell, Matiz is like the James Bond of AI poker players: adaptable, quick-thinking, and always ready for a challenge.
References: