Taxi-v3 Q-Learning — Senithu Dampegama
SYS-07 / LABORATORY SOLO — MACHINE LEARNING REINFORCEMENT LEARNING

TAXI-V3
Q-LEARNING.

A value-based reinforcement-learning agent built from scratch — no framework magic, just a Q-table, an exploration policy, and the Bellman update, trained across 10,000 episodes until pickup and drop-off became reliable.

TRAINED AGENT — PICKUP / DROP-OFF RUN
THE UPDATE RULE — APPLIED 10,000 EPISODES
Q(s,a) ← Q(s,a) + α [ r + γ · maxa′ Q(s′,a′) − Q(s,a) ]
ε-GREEDY EXPLORATION — DECAYED OVER TRAINING REWARD TRACKED PER EPISODE — CONVERGENCE MONITORED TABULAR Q OVER DISCRETE STATE × ACTION SPACE
METHOD

Epsilon-greedy exploration shifts from exploring to exploiting as epsilon decays across training. Every step applies the Bellman update to the Q-table; per-episode reward is logged to watch convergence rather than assume it.

OUTCOME + NEXT

The trained agent reliably picks up and drops off passengers. Documented next step: moving from tabular methods to Deep Q-Networks for higher-dimensional state spaces.

PYTHON NUMPY OPENAI GYM MATPLOTLIB