Interpreting Tangled Program Graphs Under Partially Observable Dota 2 Invoker Tasks
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Interpretable learning agents directly construct models that provide insight into the relationships learnt. Moreover, to date, there has been a lot of emphasis on interpreting reactive models developed for supervised learning tasks. In this work, we consider the case of models developed to address a suite of 6 partially observable tasks defined in the Dota 2 Online Battle Arena game engine. This means that learning agents need to make decisions based on the previous state as developed by the learning agent's memory; in addition to a 310-dimensional state vector provided by the game engine. Interpretability is addressed by adopting the tangled program graph approach to developing learning agents. Thus, decision-making is explicitly divide-and-conquer, with different parts of the resulting graph visited depending on the task context. We demonstrate that programs comprising the tangled program graph approach self-organize such that: (1) small subsets of task features are identified to define conditions under which index memory is written, and; (2) the subset of programs responsible for defining actions typically query indexed memory rather than task features. Particular preferences emerge for different tasks; thus, the blocking (or evasion) tasks result in a preference for specific actions whereas more open-ended tasks assume policies based on combinations of behaviours. In short, the ability to evolve the topology of the learning agent provides insights into how the policies are being constructed for addressing partially observable tasks.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it