ACT-R Model Builder

分类: 数据与AI | 上传者: HaoxuanLiTHUAI | 下载: 0 | 版本: v1.0（最新）
指导 ACT-R 认知模型构建：块类型、产生规则、子符号参数和模型验证
更新日志: Source: GitHub https://github.com/HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills
目录结构

当前层级: 根目录
📁 references/
- 📄 model-patterns.md 7.5 KB
- 📄 parameter-table.yaml 6.1 KB
📄 SKILL.md 14.3 KB
SKILL.md

---
name: "ACT-R Model Builder"
description: "Guides ACT-R cognitive model construction: chunk types, production rules, subsymbolic parameters, and model validation"
domain: "computational-cognitive-modeling"
version: "1.0.0"
papers:
 - "Anderson, 2007"
 - "Anderson & Lebiere, 1998"
 - "Anderson & Schooler, 1991"
 - "Bothell, 2023"
dependencies:
 required:
 - research-literacy
review_status: "ai-generated"
---

# ACT-R Model Builder

## Purpose

This skill encodes expert knowledge for constructing computational cognitive models within the ACT-R (Adaptive Control of Thought -- Rational) architecture. It provides guidance on chunk type definition, production rule authoring, subsymbolic parameter selection with empirically validated defaults, model fitting workflows, and validation procedures. A general-purpose programmer would not know the architecture constraints, parameter defaults, or model validation standards without specialized cognitive modeling training.

## When to Use This Skill

- Designing a new ACT-R model for a cognitive task (memory retrieval, decision-making, skill acquisition)
- Setting subsymbolic parameters and understanding their theoretical justification
- Structuring chunk types and production rules for a specific experimental paradigm
- Fitting an ACT-R model to behavioral data (RT, accuracy)
- Validating a model via parameter recovery, cross-validation, or qualitative predictions
- Choosing between ACT-R 7.x (Lisp) and pyactr (Python) for implementation

## Research Planning Protocol

Before executing the domain-specific steps below, you MUST:

1. **State the research question** -- What specific question is this analysis/paradigm addressing?
2. **Justify the method choice** -- Why is this approach appropriate? What alternatives were considered?
3. **Declare expected outcomes** -- What results would support vs. refute the hypothesis?
4. **Note assumptions and limitations** -- What does this method assume? Where could it mislead?
5. **Present the plan to the user and WAIT for confirmation** before proceeding.

For detailed methodology guidance, see the `research-literacy` skill.


## ⚠️ Verification Notice

This skill was generated by AI from academic literature. All parameters, thresholds, and citations require independent verification before use in research. If you find errors, please [open an issue](https://github.com/HaoxuanLiTHUAI/awesome_cognitive_and_neuroscience_skills/issues).

## ACT-R Architecture Overview

ACT-R is a hybrid cognitive architecture with symbolic and subsymbolic components (Anderson, 2007; Anderson & Lebiere, 1998).

### Core Modules and Buffers

| Module | Buffer | Function | Source |
|--------|--------|----------|--------|
| Declarative memory | retrieval | Stores and retrieves chunks (facts) | Anderson, 2007, Ch. 2 |
| Procedural memory | (none; fires productions) | Stores production rules (skills) | Anderson, 2007, Ch. 3 |
| Goal | goal | Tracks current task state | Anderson, 2007, Ch. 4 |
| Imaginal | imaginal | Holds intermediate problem representations | Anderson, 2007, Ch. 4 |
| Visual | visual, visual-location | Attends to and encodes visual objects | Anderson, 2007, Ch. 6 |
| Motor | manual | Executes motor responses (keypresses) | Anderson, 2007, Ch. 6 |
| Temporal | temporal | Tracks time intervals | Taatgen et al., 2007 |

### Processing Cycle

1. Buffers hold one chunk each (the "bottleneck" assumption; Anderson, 2007, Ch. 1)
2. Productions match against buffer contents (pattern matching)
3. Conflict resolution selects one production per cycle (~**50 ms** per production firing; Anderson, 2007)
4. Selected production modifies buffers or makes requests to modules
5. Modules process requests asynchronously

## Building the Symbolic Model

### Step 1: Define Chunk Types

Chunk types define the structure of declarative knowledge:

```
;; ACT-R 7.x Lisp syntax
(chunk-type addition-problem arg1 arg2 answer)
(chunk-type counting-fact number next)
```

**Decision rules for chunk type design:**

1. Each chunk type represents one category of knowledge (Anderson, 2007, Ch. 2)
2. Slots should correspond to meaningful features of the domain
3. Use inheritance when chunk types share structure (e.g., a "problem" parent type)
4. Keep chunks small -- typically **3-6 slots** per chunk type (Anderson & Lebiere, 1998)

### Step 2: Write Production Rules

Productions follow an IF-THEN structure:

```
(p retrieve-answer
 =goal>
 isa addition-problem
 arg1 =num1
 arg2 =num2
 answer nil
 ?retrieval>
 state free
==>
 +retrieval>
 isa addition-fact
 addend1 =num1
 addend2 =num2
 =goal>
)
```

**Production rule guidelines:**

| Guideline | Rationale | Source |
|-----------|-----------|--------|
| One request per production | Module bottleneck constraint | Anderson, 2007, Ch. 3 |
| Test buffer state before requesting | Prevents jamming the module | Bothell, 2023, ACT-R reference manual |
| Use `=goal>` to maintain goal buffer | Prevents goal harvesting | Bothell, 2023 |
| Minimize productions per task step | Simpler models are preferred (parsimony) | Anderson, 2007, Ch. 1 |

### Step 3: Structure the Goal Stack

```
Is the task sequential with clear phases?
 |
 +-- YES --> Use a single goal chunk with a "step" slot
 | that tracks the current phase
 |
 +-- NO --> Does the task require subgoaling?
 |
 +-- YES --> Use goal push/pop (stack)
 |
 +-- NO --> Use the imaginal buffer for
 intermediate representations
```

## Subsymbolic Parameters

These parameters govern memory activation, retrieval, and production selection. See `references/parameter-table.yaml` for the complete table.

### Core Declarative Memory Parameters

| Parameter | Symbol | Default | Typical Range | Source |
|-----------|--------|---------|---------------|--------|
| Base-level learning decay | d | **0.5** | 0.1 -- 1.0 | Anderson & Schooler, 1991; Anderson, 2007 |
| Activation noise | s | **0.4** | 0.1 -- 0.8 | Anderson, 2007 |
| Latency factor | F | **1.0** | 0.2 -- 5.0 | Anderson, 2007 |
| Latency exponent | f | **1.0** | Fixed in most models | Anderson, 2007 |
| Retrieval threshold | tau | **-infinity** (default) | Set empirically; often **0.0** to **-2.0** | Anderson, 2007 |
| Maximum associative strength | S (mas) | **context-dependent** | 1.0 -- 5.0 | Anderson & Reder, 1999 |
| Mismatch penalty | P | **application-dependent** | 0.5 -- 2.0 | Anderson, 2007 |

### Production Utility Parameters

| Parameter | Symbol | Default | Typical Range | Source |
|-----------|--------|---------|---------------|--------|
| Utility noise | sigma | **0.0** (deterministic) | 0.1 -- 2.0 when enabled | Anderson, 2007 |
| Utility learning rate | alpha | **0.2** | 0.01 -- 1.0 | Anderson, 2007 |
| Initial utility | U0 | **0.0** | Set per production | Anderson, 2007 |
| Production compilation | enabled/disabled | Disabled by default | -- | Taatgen & Anderson, 2002 |

### Timing Parameters

| Parameter | Value | Source |
|-----------|-------|--------|
| Production cycle time | **50 ms** | Anderson, 2007 |
| Visual encoding time | **85 ms** | Anderson, 2007, Ch. 6 |
| Motor initiation time | **50 ms** | Anderson, 2007, Ch. 6 |
| Motor execution time | **100 ms** (Fitts' law applies) | Anderson, 2007, Ch. 6 |
| Imaginal delay | **200 ms** | Anderson, 2007, Ch. 4 |

### Activation Equation

Total activation of chunk i:

**A_i = B_i + sum_j(W_j * S_ji) + PM_i + noise**

Where:
- B_i = base-level activation (log of weighted recency; decay d; Anderson & Schooler, 1991)
- W_j * S_ji = spreading activation from source j (Anderson, 2007, Ch. 5)
- PM_i = partial matching component (Anderson, 2007)
- noise = logistic noise with scale s (Anderson, 2007)

Retrieval time: **RT = F * e^(-f * A_i)** (Anderson, 2007)

## Model Fitting Workflow

### Step 1: Identify Free Parameters

```
How many free parameters?
 |
 +-- <= 3 --> Standard practice; proceed
 |
 +-- 4-6 --> Acceptable if justified by model complexity
 |
 +-- > 6 --> Warning: overfitting risk. Consider fixing some
 to default values (Anderson, 2007)
```

**Rule of thumb**: The number of free parameters should be substantially less than the number of independent data points being fit (Roberts & Pashler, 2000).

### Step 2: Choose Fitting Method

| Method | When to Use | Source |
|--------|-------------|--------|
| Grid search | Few parameters (1-3), bounded space | Standard practice |
| Simplex (Nelder-Mead) | Moderate parameters, smooth landscape | Anderson, 2007 |
| Differential evolution | Many parameters, multimodal landscape | Storn & Price, 1997 |
| Bayesian optimization | Expensive evaluations, informed priors | Palestro et al., 2018 |

### Step 3: Fit to Multiple Dependent Variables

ACT-R models should simultaneously account for:
- **Response times** (correct trials, mean or quantiles)
- **Accuracy** (proportion correct by condition)
- **Qualitative patterns** (error types, learning curves)

Use weighted sum of squared deviations or log-likelihood across measures (Anderson, 2007, Ch. 4).

### Step 4: Parameter Recovery

Before trusting fitted parameter values, conduct a parameter recovery study. See the `parameter-recovery-checker` skill.

## Common Model Patterns

See `references/model-patterns.md` for detailed implementations of:

1. **Memory retrieval** -- Paired associates, fan effect (Anderson, 2007, Ch. 5)
2. **Skill acquisition** -- Production compilation, power law of practice (Taatgen & Anderson, 2002)
3. **Decision-making** -- Instance-based learning, utility-based selection (Gonzalez et al., 2003)
4. **Problem solving** -- Means-ends analysis, goal stacking (Anderson, 2007, Ch. 8)

## Model Validation Checklist

| Validation Step | Method | Minimum Standard |
|----------------|--------|-----------------|
| Parameter recovery | Simulate and refit | r > **0.9** between true and recovered (Heathcote et al., 2015) |
| Cross-validation | Fit half, predict half | Prediction RMSE within **2x** of fitting RMSE |
| Qualitative predictions | Novel conditions | Model predicts ordinal pattern correctly |
| Model comparison | AIC/BIC or Bayes factor | Compare against plausible alternatives (Burnham & Anderson, 2002) |
| Sensitivity analysis | Vary fixed parameters | Conclusions robust to +/-**20%** variation |

## Software and Implementation

| Platform | Language | URL | Notes |
|----------|----------|-----|-------|
| ACT-R 7.x | Common Lisp | act-r.psy.cmu.edu | Reference implementation (Bothell, 2023) |
| pyactr | Python | github.com/jakdot/pyactr | Python interface, good for batch simulations (Dotlacil, 2018) |
| jACT-R | Java | jactr.org | Java implementation |

**Recommendation**: Use ACT-R 7.x for model development and validation. Use pyactr when integrating with Python data analysis pipelines or running large parameter sweeps (Dotlacil, 2018).

## Common Pitfalls

1. **Too many free parameters**: Fitting more than **5-6** free parameters without strong justification risks overfitting (Roberts & Pashler, 2000). Fix well-established parameters (d = **0.5**, production cycle = **50 ms**) to defaults.
2. **Ignoring parameter correlations**: Parameters like s (noise) and tau (threshold) trade off. Run parameter recovery to verify identifiability.
3. **Fitting only means**: ACT-R makes distributional predictions. Fitting only mean RT discards information. Use quantile-based fitting where possible (Heathcote et al., 2015).
4. **Incorrect timing alignment**: ACT-R's predicted RT includes perceptual and motor times. Account for these when comparing to behavioral RT.
5. **Overly complex models**: Prefer models with fewer productions and chunk types that still capture the qualitative pattern. Complexity should be motivated by the data (Anderson, 2007, Ch. 1).
6. **Neglecting model comparison**: Always compare your ACT-R model against at least one alternative (simpler ACT-R variant or a different architecture) using formal model comparison (Burnham & Anderson, 2002).

## Minimum Reporting Checklist

Based on best practices from Anderson (2007) and Heathcote et al. (2015):

- [ ] Architecture version (e.g., ACT-R 7.27)
- [ ] All chunk types and their slots
- [ ] Number of production rules
- [ ] All parameter values: fixed (with default source) and free (with fitted values and confidence intervals)
- [ ] Fitting method and objective function
- [ ] Data fitted: number of conditions, number of data points, dependent variables
- [ ] Goodness of fit: R-squared, RMSE, or log-likelihood per dependent variable
- [ ] Parameter recovery results (r, bias, RMSE for each free parameter)
- [ ] Model comparison results (AIC/BIC/Bayes factor vs. alternatives)
- [ ] Qualitative predictions and whether they matched data

## References

- Anderson, J. R. (2007). *How Can the Human Mind Occur in the Physical Universe?* Oxford University Press.
- Anderson, J. R., & Lebiere, C. (1998). *The Atomic Components of Thought.* Lawrence Erlbaum Associates.
- Anderson, J. R., & Reder, L. M. (1999). The fan effect: New results and new theories. *Journal of Experimental Psychology: General*, 128(2), 186-197.
- Anderson, J. R., & Schooler, L. J. (1991). Reflections of the environment in memory. *Psychological Science*, 2(6), 396-408.
- Bothell, D. (2023). *ACT-R 7.27 Reference Manual.* Carnegie Mellon University.
- Burnham, K. P., & Anderson, D. R. (2002). *Model Selection and Multimodel Inference* (2nd ed.). Springer.
- Dotlacil, J. (2018). Building an ACT-R reader for eye-tracking corpus data. *Topics in Cognitive Science*, 10(1), 144-160.
- Gonzalez, C., Lerch, J. F., & Lebiere, C. (2003). Instance-based learning in dynamic decision making. *Cognitive Science*, 27(4), 591-635.
- Heathcote, A., Brown, S. D., & Wagenmakers, E.-J. (2015). An introduction to good practices in cognitive modeling. In B. U. Forstmann & E.-J. Wagenmakers (Eds.), *An Introduction to Model-Based Cognitive Neuroscience*. Springer.
- Palestro, J. J., Sederberg, P. B., Osth, A. F., Van Zandt, T., & Turner, B. M. (2018). Likelihood-free methods for cognitive science. Springer.
- Roberts, S., & Pashler, H. (2000). How persuasive is a good fit? A comment on theory testing. *Psychological Review*, 107(2), 358-367.
- Storn, R., & Price, K. (1997). Differential evolution. *Journal of Global Optimization*, 11(4), 341-359.
- Taatgen, N. A., & Anderson, J. R. (2002). Why do children learn to say "Broke"? A model of learning the past tense without feedback. *Cognition*, 86(2), 123-155.
- Taatgen, N. A., van Rijn, H., & Anderson, J. R. (2007). An integrated theory of prospective time interval estimation. *Psychological Review*, 114(3), 577-598.

See `references/` for detailed parameter tables and common model patterns.
登录后下载/点赞/收藏 ❤ 14 | ★ 0
ACT-R Model Builder

目录结构

SKILL.md

举报内容

提示
