Navigate to a goal
- After its training phase
- Not yet measured
- End of stream
- Not yet measured
- Learning curve
- Not yet measured
Clark Whitehead Research
Research progress
Learn new tasks while preserving previously learned capabilities.
Planning — not yet evaluated
No results have been measured for this proposed evaluation. The environment, budget, and targets are provisional; pilot checks will inform a versioned protocol before formal evaluation. Architecture choices remain open.
The first diagnostic
Custom four-task navigation diagnostic. A controlled, custom diagnostic to measure learning and forgetting cheaply before broader agent evaluations.
Provisional custom diagnostic, not an established benchmark. It is an initial check of learning and forgetting, not the final scope of the research. Success would justify broader testing, not demonstrate general continual learning or AGI.
Toward survival across worlds
The broader proposal is to test survival as a primary reward in changing multiagent worlds, then study continual adaptation across environments while retaining earlier capabilities. Begin with small, controlled populations. Millions of interacting agents are a long-term ambition, not a current system or measured achievement.
The gridworld plan below isolates basic learning and forgetting before that broader program. Its existing numerical targets remain provisional and apply only to this diagnostic; survival and multiagent evaluation criteria have not yet been defined. Explore the survival hypothesis and its open questions.
Environment & comparisonsCustom diagnostic · provisional budget
Four procedurally generated gridworld task families share the same fully observed interface and actions, with visible task cues and sequential training. Proposed layouts are 9 by 9, with a proposed 100-action episode limit.
- Navigate to a goal
- Collect a key, then reach a locked goal
- Avoid hazards while reaching a goal
- Combine keys, locks, and hazards
Main comparisons
- Sequential training without replay. The forgetting baseline.
- Fixed-capacity uniform replay. A memory-based comparison with disclosed capacity and contents.
- Candidate method. The approach to evaluate; architecture choices remain open.
Proposed budget
100,000 environment interactions per task. Match environment-interaction and gradient-update budgets across the main comparisons, and count replay updates.
Disclose model size, memory capacity and contents, tuning effort, runtime and compute.
Optional joint-training and task-specialist references would be labeled separately because they have different data-access assumptions.
This is a provisional custom diagnostic, not an established benchmark or a commitment to a game-focused research program. Success would justify broader testing, not demonstrate general continual learning or AGI.
Proposed targets & measurementsEvery outcome: Not yet measured
Proposed targets only — not achieved scores or calibrated promises. Targets remain subject to pilot calibration.
| Proposed target | Measured outcome |
|---|---|
| Immediate new-task successAt least 70% held-out success on each task immediately after training it. | Not yet measured |
| Average retention lossAt most 10 percentage points on earlier tasks at the end of the stream. | Not yet measured |
| Final average vs. sequential no-replayAt least 10 percentage points higher final average success. | Not yet measured |
| Immediate new-task average vs. replayNo more than 5 percentage points below the replay baseline. | Not yet measured |
| Final average vs. replayReport the comparison. Do not claim an improvement over replay unless evidence supports it.Reporting measure; no assumed win. | Not yet measured |
Per-task outcomes
Held-out success immediately after training each task, and again at the end of the stream.
Collect a key, then reach a locked goal
- After its training phase
- Not yet measured
- End of stream
- Not yet measured
- Learning curve
- Not yet measured
Avoid hazards while reaching a goal
- After its training phase
- Not yet measured
- End of stream
- Not yet measured
- Learning curve
- Not yet measured
Combine keys, locks, and hazards
- After its training phase
- Not yet measured
- End of stream
- Not yet measured
- Learning curve
- Not yet measured
Methods & safeguardsMeasurement definitions, fairness, and scope
New-task success
Measure success on held-out layouts immediately after each task phase.
Retention loss
For each of the first three tasks, calculate max(0, immediate success − end-of-stream success), then take the mean. Improvements on one task cannot conceal forgetting on another.
Full reporting
Report per-task final success and learning curves alongside aggregate metrics, including final average success versus replay.
Evaluation safeguards
- Use separate training, validation, and final-test generation seeds, with duplicate-layout checks.
- Freeze the task order, hyperparameters, budgets, thresholds, and evaluator before final evaluation.
- Allow no model updates or replay writes from test episodes.
Seeds, task orders & uncertainty
Use at least 3 independent training seeds for initial preliminary evidence and report all of them. Later validation targets 10 seeds over 3 predeclared task orders, with uncertainty intervals.
Pilot calibration before freezing
The numbers are proposed targets, not achieved scores or calibrated promises. If pilot baselines show ceiling effects or little forgetting, revise the diagnostic using development data before freezing it, rather than claim success on an uninformative test.
Outside this initial diagnostic
Hidden task boundaries, Partial observability, Real-world transfer, General agent ability. Future established benchmark selection remains pending.
Research milestonesPlanning and planned; no completed research milestones
- planning
Draft and freeze protocol
Draft proposed; freeze after pilot calibration and before formal evaluation.
- planned
Implement and test environment
- planned
Establish sequential and replay baselines
- planned
Evaluate candidate
- planned
Repeat across seeds and task orders
- planned
Evaluate on an independently selected established benchmark
Benchmark selection remains pending.
Results & dated progress logInitial proposed plan only
Results
No measured results yet for this proposed evaluation.
Future reports will include baseline comparisons, resource costs, variability, failures, and limitations.
- Resource costs
- Not yet measured
- Variability & uncertainty
- Not yet measured
Progress log
Research direction clarified — planning only
Survival rewards and continual adaptation across simulated multiagent worlds are proposed research. Millions of agents remain a long-term ambition. No experiments or measured results are being reported; the existing diagnostic and its provisional targets are unchanged.
Next step: Review reward comparisons and evaluation controls, begin with small controlled populations, and calibrate the initial learning-and-forgetting diagnostic before freezing a protocol.
Proposed plan published
Plan only. No measured results for this proposed evaluation; the protocol remains provisional.
Next step: Review the draft and calibrate the diagnostic with development data before freezing a protocol for formal evaluation.
Progress updates are manually reviewed and published. This page does not run experiments or collect measurements automatically.