F1
The plan cannot see the road
No sensing, no feedback. The controller runs the same schedule into an empty street and a saturated one.
ATCS-GH case study
Pitch deck
ATCS-GHAdaptive Traffic Control System — Ghana
A deep reinforcement learning controller that replaces fixed-timer signal plans — reading each junction in real time and choosing its next phase every 5 seconds.
Median wait reduction
86%
vs. fixed-time control, across 8 scenarios
Scenarios improved
8 / 8
and at every random seed, not only on average
Status
Simulation-validated · Pre-pilot
No field deployment yet. All results are from simulation.
Founder
Osborn Dogbe
Valiborn Technologies · Accra, Ghana
02 · The problem
Ghana loses GH¢4.5 billion a year to congestion.
The cost is not abstract. It is measured hours, burnt fuel, and degraded working capacity — concentrated in the same Accra corridors every weekday morning and evening.
Annual cost breakdown
- Time lost sitting in trafficGH¢3.2 bn · 71%
- Productivity lost to congestion-related stress and fatigueGH¢815 m · 18%
- Wasted fuelGH¢434 m · 10%
Source: “The Cost of Gridlock: A Policy Brief on Ghana’s Urban Traffic Crisis,” Glima Research, February 2026.
03 · Why fixed timers fail
A green stays green for a preset number of seconds — whether twenty cars are queued, or none.
F2
Surges are invisible
A market-day surge or an incident produces demand the fixed plan has no mechanism to absorb.
F3
Rush hour flips direction
Morning flow is inbound, evening outbound. One split cannot serve both, so one peak is always mistimed.
F4
Junctions don't coordinate
Traffic cleared at one light simply queues again at the next 300 m down the corridor.
04 · The solution
Replace the schedule, not the hardware.
ATCS-GH is a software controller. It keeps the existing poles, heads and phase order, and substitutes the one component that is failing: the decision about which phase runs next, and for how long.
01
Software-first
Runs end-to-end on a laptop. No proprietary controller cabinet, no vendor lock.
02
Corridor-aware
Three independent agents, one per junction, with no explicit communication — each sees its neighbours through its own state vector.
03
Built for local mixed traffic
Private cars, shared trotros and signal-respecting pedestrian crossings — modelled on real Accra geometry, not a generic grid.
Test bed — N6 Nsawam corridor
Three signalised junctions 300 m apart — Achimota/Neoplan, Asylum Down, Nima/Tesano — plus a single-junction bed at 5.6216°N 0.2193°W. Three independent agents, no explicit communication. Both beds are calibrated from real Accra geometry in SUMO, the microsimulator used by transport agencies.
05 · How it works
Sense → decide → act, every five seconds.
Sense
41 / 46 dims41 at a single junction; 46 per junction on the corridor.
- Per-lane queue length
- Vehicle speed
- Accumulated waiting time
- Corridor: neighbour queue, neighbour phase, link occupancy
Decide
Double-DQNPyTorch. One agent per junction — the corridor runs three, trained independently.
- Value estimate per phase
- Bounded and clipped reward
- No scripted override of any kind
Act
5 s intervalOne of five protected-left phases is committed to the signal, then the loop closes.
- SUMO applies the phase
- Godot 4 renders it live over WebSocket
- Flask dashboard logs the metrics
Loop closes — 720 decisions per hour, per junction
06 · Action space & safety
Five phases. Protected left only.
0
HOLD
Keep the current phase running — extend a green that is still clearing queue.
1
NS_THROUGH
North–south straight movements.
2
NS_LEFT
North–south protected left turns, no conflicting through movement.
3
EW_THROUGH
East–west straight movements.
4
EW_LEFT
East–west protected left turns, no conflicting through movement.
Deliberate exclusion
Permissive and all-green phases are removed from the action space. They deadlock the junction box: opposing streams enter the intersection together, block each other mid-turn, and no subsequent phase can clear them. This was not a design assumption — it was a real failure observed in training, diagnosed, and fixed by constraining the action space. A hard-coded emergency-vehicle preemption rule was also removed, so every result here is purely learned behaviour with no scripted override propping it up.
07 · Training methodology
The differentiator is not the network. It is how it is trained and selected.
Deep RL on traffic is well-trodden. What makes these results hold up is a set of decisions aimed squarely at the failure modes that usually break it: reward blow-up under saturation, useless early exploration, scenario overfitting, and best-case checkpoint selection.
T1
Bounded, clipped rewards
Saturated traffic cannot produce a runaway reward signal. Without the bound, a gridlocked junction generates penalties large enough to trap the agent in the gridlock that caused them.
T2
Expert warm-start
70% of exploratory actions follow a sustained-green heuristic. Heavy scenarios keep flowing while exploration decays, so the agent learns from traffic that moves instead of from a car park.
T3
Scenario rotation
One demand scenario per episode, round-robin. The agent never gets to specialise in the easy case — every scenario keeps returning until it is handled.
T4
Maximin model selection
The shipped checkpoint is the one whose worst scenario — or worst junction, on the corridor — performs best. Not its best case. Gated to near-greedy exploration, so selection reflects deployment behaviour.
08 · Results
86% median reduction in mean average vehicle waiting time.
Every number here comes from a greedy policy with exploration switched off, run over full two-hour scenarios, five random seeds each. Nothing is a training-time best case.
8 / 8
Scenarios improved
No scenario where the learned controller is worse.
40
Evaluation runs
8 scenarios × 5 seeds, greedy policy, ε = 0.
7200 s
Per run
Full two-hour demand profiles, not short episodes.
≈ 2 h
To train the corridor
Three agents, on a laptop. No cluster required.
| System | Scenario | ATCS-GH | Fixed timer | Reduction |
|---|---|---|---|---|
| Corridor | Evening rush | 10.5 s | 134.8 s | 92% |
| Corridor | Morning rush | 13.0 s | 143.1 s | 91% |
| Corridor | Off-peak | 3.4 s | 25.6 s | 87% |
| Single | Weekend market✳ | 13.8 s | 885.8 s | 98% |
| Single | Morning rush | 27.8 s | 186.7 s | 85% |
| Single | Continuous day | 18.1 s | 47.6 s | 62% |
| Single | Off-peak | 8.4 s | 14.2 s | 41% |
| Single | Evening rush | 256.1 s | 427.4 s | 40% |
Mean average vehicle waiting time, seconds — lower is better. Greedy policy, ε = 0, 7200 s runs, 5 seeds each. Baseline is a realistic protected-left fixed timer: NS 40 s through / 15 s left, EW 25 s / 10 s. ✳ collapse-avoidance, not typical performance — see section 09.
09 · How to read these numbers
What these results are not.
A reviewer's first instinct on seeing a 98% improvement should be suspicion. So here is the reading we would give ourselves, unprompted — the honest headline is 85–92% on peak-hour scenarios, and every qualifier below is load-bearing.
C1The 98% figure
98% is collapse-avoidance, not typical performance.
In the weekend-market scenario the fixed-timer baseline is a gridlocked junction, so the agent is preventing a breakdown rather than improving a plan that works. Read it as robustness, not speed-up. The honest headline is 85–92% on peak-hour scenarios.
C2The comparison
The comparison is always against fixed-time control.
That is the status quo across much of Accra, which is why it is the right baseline. Against a modern vehicle-actuated system the margin would be smaller — we have not measured that comparison and do not claim it.
C3The setting
These are simulation results. There is no field deployment.
Status is simulation-validated, pre-pilot. Everything reported here comes from SUMO with calibrated Accra geometry. No vehicle in Accra has yet been stopped or released by this controller.
C4The sensing
The agent currently reads idealised sensor data.
Real cameras and inductive loops give noisy, dropout-prone readings. Field performance is expected to be lower than these figures, and quantifying that gap is the next phase of work, not an afterthought.
C5The demand
Demand was calibrated to solvable levels.
Both controllers are graded against an achievable target rather than an impossible one. Loading the network past the point where any signal plan can cope would inflate the relative improvement without saying anything useful.
10 · Validation rigour
Four choices that made the results harder to get.
V1
An honest baseline
A realistic protected-left fixed timer — NS 40/15 s, EW 25/10 s. An earlier, easier baseline was discarded for being too easy to beat.
V2
Worst-case selection
Maximin checkpointing: we ship the model with the best worst case, so the reported numbers are not a lucky snapshot cherry-picked from a training curve.
V3
Deployment conditions
Exploration off (ε = 0), full 7200 s scenarios, 5 seeds each, 40 runs total. Evaluated the way it would run on the street, not the way it trains.
V4
Reproducibility
Fixed seeds, a scripted evaluation harness, and the whole pipeline runs on a laptop — so any reviewer can re-run it rather than take our word for it.
11 · What's built
A working platform, not a notebook.
Five components exist and run today, end to end, on commodity hardware — simulation, agents, live 3D visualisation, analytics, and the evaluation harness that produced every figure above.
01SUMO
Calibrated SUMO networks
Achimota/Neoplan junction and the 3-junction N6 Nsawam corridor, built on real geometry with mixed traffic — private cars, shared trotros, signal-respecting pedestrian crossings.
02PyTorch
Trained Double-DQN agents
One agent per junction; three independent agents on the corridor with no explicit communication. Shipped checkpoints selected by worst-case performance.
03Godot 4
Real-time 3D visualiser
A full Accra streetscape — vehicles, pedestrians, day/night cycle and weather — driven live from the simulation over WebSocket. Built for non-technical stakeholders, including city officials.
04Flask
Analytics dashboard
Live queue, wait and throughput telemetry per junction and per phase, so behaviour can be inspected rather than inferred from a summary number.
05Python
Evaluation harness
Scripted multi-seed, multi-scenario benchmarking with exploration disabled — the harness that produced all 40 runs, re-runnable end to end on a laptop.
12 · Transferability & scaling
We tested whether a trained policy transfers. It does not.
A swap test dropped each junction's trained policy onto the other junctions with zero retraining. Every swap failed outright. Lane layouts differ, so the inputs fall outside the model's training distribution — and on a corridor the damage compounds, because one misbehaving junction backs up into its neighbours.
Swap test — trained policy → target junction
J0 → J1 FAILEDJ0 → J2 FAILEDJ1 → J0 FAILEDJ1 → J2 FAILEDJ2 → J0 FAILEDJ2 → J1 FAILED
Zero retraining, greedy policy. Reported because it is informative, not because it flatters the system.
The consequence, stated plainly
The product is not a checkpoint. It is the pipeline that produces one.
Network calibration, reward and training design, and the evaluation harness are the durable assets. A new junction is a retraining job measured in hours on commodity hardware, warm-started from existing weights.
13 · Roadmap
From simulation to a supervised live corridor.
01Complete
Simulation validation
Multi-seed benchmarking across 8 scenarios, the 3D visualiser, and a reproducible evaluation harness are all shipped and running.
02Next
Realistic perception
Camera-based vehicle and queue detection, injected sensor noise and dropout, and an explicit measurement of how much performance degrades under it.
03Then
Shadow-mode pilot
Pilot engagement with a metropolitan road authority, instrumenting one live junction in shadow mode — recommending phases without controlling them.
04Then
Supervised live control
Supervised control of a pilot corridor with measured before/after outcomes, extended junction by junction as each one clears its own evaluation.
Shadow mode means the controller recommends phases and is measured against the live timer without ever driving the signal. Nothing touches public traffic until it has been graded that way first.
14 · Why Google Africa Applied AI Lab
The controller works. What's missing is perception and scale.
Between this deck and a live Accra junction sit two hard problems and one open research question. All three are things the Lab is unusually well placed to unblock.
01Highest leverage
Vision models for detection
Inductive loop sensors are expensive to install and maintain. Robust vehicle and queue estimation from ordinary roadside cameras is the single change that makes this deployable at African cost structures — and it is the gap between a validated controller and a live junction.
02Timeline
Compute for network-scale training
Three coordinated agents train on a laptop. Tens of junctions across a district is a different problem — more agents, more interaction, far more scenario coverage. Serious training capacity moves that from years to months.
03Open questions
RL expertise
Multi-agent stability, sim-to-real transfer, and safe deployment of a learned policy onto public infrastructure are open questions. They are answered better with Google engineers than alone.
04Reach
A named African deployment case
A measured before/after on a real Accra corridor generalises directly to Lagos, Nairobi and Kampala — the same fixed-timer infrastructure, the same congestion economics, the same absent capital budget for new sensor hardware.
15 · Contact
Osborn Dogbe
Founder, Valiborn Technologies · Accra, Ghana
ideally47@gmail.comSimulation-validated · Pre-pilot · Results are simulation-derived, no field deployment