Glonce

Game of Hidden Rules report trains RL agents to infer rules by trial and error

researchThe Glonce editorsarXiv🔥66Glonce rating66Impact60Novelty65Relevance95Surprise35Players25Credibility85

The report covers work on the Game of Hidden Rules, or GOHR: a setting where reinforcement learning agents must work out a rule that governs the game without being told what it is, learning instead from trial-and-error feedback on their moves. The agents are built on a Transformer-based A2C framework, meaning the advantage actor-critic reinforcement learning algorithm paired with a Transformer network as the underlying model.

A central strand of the work is representation design: the report compares a Feature-Centric representation against an Object-Centric representation as two different ways of encoding the game state for the agent, and examines how that choice shapes learning. Alongside this, the report analyzes how the difficulty of a hidden rule affects how well agents learn it, and studies transfer learning and generalization: whether an agent that has learned to infer one set of hidden rules can carry that skill over to new, unseen rules rather than starting from scratch each time.

The report also extends beyond the RL agents themselves to pseudo-bot-assisted analysis of human learning data, classifying how people go about inferring the same hidden rules in the game. This puts machine rule-inference and human rule-inference side by side within the same GOHR framework.

The text summarized here is scoped as a project report describing this line of work; it does not itself state the authors, their institutional affiliation, a publication venue, or a submission date, nor does it give the accuracy figures, scores or sample sizes produced by the experiments, or how the results compare against prior methods.

Key facts

  • The report trains reinforcement learning agents to infer hidden game rules in the Game of Hidden Rules (GOHR) benchmark purely from trial-and-error feedback.
  • The agents use a Transformer-based A2C (advantage actor-critic) framework.
  • It compares two ways of representing the game state to the agent: Feature-Centric and Object-Centric representations.
  • It studies how rule difficulty affects learning, and how well agents transfer and generalize their rule-inference skill across different rule sets.
  • It also analyzes human learning data with pseudo-bot assistance, classifying how people infer the same hidden rules.

Why it matters

Rule inference from trial-and-error feedback is a stand-in for a broader problem in reinforcement learning: getting an agent to work out an unstated structure in its environment rather than being handed the objective directly. Studying this inside a controlled benchmark like GOHR, with two competing state representations and an explicit transfer-learning test, gives a structured way to ask which design choices actually help an agent generalize its rule-inference skill instead of memorizing one specific rule set. Pairing that with an analysis of human rule-learning data in the same framework is a further step: it lets machine and human strategies for the same task be examined side by side rather than only in isolation.

Who it affects

The most direct audience is reinforcement learning researchers working on generalization, transfer learning and representation design, since the report's comparisons speak to those choices specifically. It also bears on researchers studying human concept learning and rule inference, given the human-learning-data component, and on anyone building or evaluating rule-inference benchmarks, since GOHR itself is the testbed the report works within.

How to use it

This is a research report describing experimental work and methodology, not a released product, tool or dataset with stated pricing or licence terms; none are given in the text covered here. The practical takeaway for a reader working in this area is the design choices themselves: the Transformer-based A2C setup and the Feature-Centric versus Object-Centric representation comparison offer a concrete pair of options to weigh when building an agent meant to infer structure from feedback rather than from an explicit specification.

How solid is it

The text summarized here describes the report's scope and structure but does not carry the experimental results: no accuracy figures, scores or sample sizes for either the RL agents or the human-learning analysis are given, and no comparison against prior methods or baselines is stated. It also does not name the authors, their institution, a publication venue or a submission date. Those gaps sit in what was available to summarize, not necessarily in the underlying report itself, so a reader who wants the actual findings and their statistical weight needs the full report.

Risks and caveats

Because the material available here is a scope-level summary rather than the results section, the specific claims about how well representation choice, transfer learning or human-comparison analysis actually worked cannot be verified from this text alone. The definition of what a 'hidden rule' concretely looks like inside GOHR is also not spelled out in what was summarized, which limits how precisely the setup can be judged from this account.