Dealer —

The agent conditions on the upcard only.

Agent —

Deal a hand to start.

What the agent sees
(–,–,–)
player sumdealer upcardusable ace
Action values
hit
—
stand
—

Waiting for a hand.

Decisions by
0hands
0won
0lost
0pushed
0.00net units

Learned policy

Every state the agent can be in, coloured by the action it would take. Saturation is Q(hit) − Q(stand) — how sure it is.

Dealer upcard Player sum
stand
hit
never visited (heuristic) differs from basic strategy
Table view

Train

Episodes run in the C++ server on a worker thread. The grid on the right is the live policy — watch it resolve out of noise.

0episodes run
—avg reward, last window
—win rate, last window
—epsilon
—states learned

Average reward

units per hand — the metric that matters

Win rate

wins ÷ decided hands · capped near 44% by the rules

Exploration

epsilon — probability of a random action

episodes

Policy, live

hard totals

Start training to watch the table fill in.

Q-learning vs Monte Carlo

Same rules, same deck, two update rules. Both play greedily.

No run yet. Deal a batch to compare the two agents.