Teaching AI to Bet – The Q-Learning Roulette Agent

Can an agent learn how to bet more intelligently than a fixed system?

This post is part of a multi-part series on simulating and analyzing roulette betting strategies with Python and reinforcement learning.

• Start from the beginning: The House Always Wins? I Taught 3 AI Agents to Beat Roulette Anyway
• Explore the code: GitHub – rossautomatedsolutions/roulette-simulator

---

In earlier posts, I analyzed classical systems like Flat Betting and Martingale. They all failed in the long run.

Now, I explored something more dynamic: a reinforcement learning agent that learns how to bet based on feedback from wins and losses.

This was my first “intelligent” system — built using tabular Q-learning.

---

What Is Q-Learning?

Q-learning is a reinforcement learning algorithm that learns the expected reward of taking an action in a given state. Over time, the agent builds a Q-table that maps:

(State, Action) → Value

The agent starts by exploring randomly and eventually settles into the most profitable decisions.

---

My Approach

I created a custom RouletteEnv environment with a small, discrete state space:

• Bankroll (bucketed by $100)
• Previous spin result (win/loss)
• Last outcome (bucketed into 6 bins)

Action space included outside bets (red, black, even, etc.) and straight number bets (0–36). Each action bet a flat $20.

Reward shaping included:

• Payout from win/loss
• Bonus for surviving 100+ spins
• Penalty for going bust
• Bonus for exceeding 1.5x bankroll

---

Training Results (1000 Episodes)

During training:

• The agent learned to favor safer bets early (e.g., red, black)
• It avoided large losing streaks more effectively than classic systems
• It sometimes found profitable sequences using number-based bets

---

Sample Evaluation Results (100 Runs)

Strategy	Avg Reward	Median Reward	Avg Spins	% Profitable
Q-Learning Agent	$8,522	-$1,180	375	33%
Flat Betting	-$626	-$1,120	775	8%

While the agent didn’t consistently profit, it had higher upside potential than any fixed strategy I tested. It learned to take calculated risks — occasionally producing 5-digit returns — but also suffered busts.

---

Takeaway

Q-learning produced mixed results. It didn’t beat the house consistently, but it clearly learned non-random behavior.

• The agent outperformed Flat Betting in 33% of runs
• It was more volatile — but not completely reckless
• Reward shaping and state design had a big impact on behavior

This opened the door to more nuanced training techniques and smarter agents. In the next post, I added risk-awareness to encourage survival alongside reward.

---

Continue to Post 4: Smarter ≠ Safer – Risk-Aware Q-Learning