Skip to content

Five games. Learning, search and knowing when to stop.

Explore trained players and classical puzzle solvers: how each works, what the evidence shows, and when further training is unnecessary.

Hex-a-Hop

Can classical planning finish every level, and what help did the last one need?

Status: Complete route coverage · 99 independent + 1 assisted

100 / 100 verified routes99 independently searched; Fetch Quest solved with published walkthrough assistance. No trained campaign model.

Explore the Hex-a-Hop experiment →

Sokoban

When is a search algorithm a better first choice than machine learning?

137 / 137 puzzles solvedClassical search, zero training runs. Why a verified algorithm was enough for this community campaign.

Explore the Sokoban experiment →

Snake

Can a model fill every cell and still take useful shortcuts?

120 / 120 full-board wins40.06% fewer movements on 60 fresh starts. Watch verified recordings of three courses.

Explore the Snake experiment →

Connect Four

How much playing strength comes from learning, and how much from search?

87.5% against our ExpertLocal network + native search controller, 100 games, draws half credit. Raw policy is much weaker; browser ML play is pending.

Explore the Connect Four experiment →

Freedoom

Can a neural player finish a 3D level using only pixels and memory?

Status: Research paused · reliable-player goal unmet

340 / 400 level completions85% on familiar MAP01, 0/40 on unfamiliar MAP02. Research paused; the reliable-player goal remains unmet. This is a MAP01 specialist, not a completed campaign.

Explore the Freedoom experiment →

What the five case studies taught us

Define success in the game before choosing a training recipe. Snake needs to fill the whole board; Connect Four needs strong moves at a fair thinking budget; Freedoom needs real level exits. Sokoban first needs a complete, verified puzzle campaign; exact search already solves this collection. Hex-a-Hop has 100 verified routes, with published assistance for the final unresolved puzzle. Lower loss or more collected rewards can help diagnose training, but neither replaces those outcomes.

Describe the whole player honestly. Snake uses engineered features and expert demonstrations. Connect Four gains much of its strength from search. The selected Freedoom model learns from demonstrations and plays with pixels and memory, but has not met its reliability goal. Separate learning, planning and assistance so each result means something.

Keep fresh evaluation separate from fitting, retain failures, and stop when the next run has no specific hypothesis or budget. A clear, verified demonstration can be a useful learning product even when the research target remains unmet.

Each page brings the training story, explanations, results and source reports together. Terms are explained where you first need them.