Abstract
Ideomotor Theory (IMT) suggests that goal-directed behavior is driven by the anticipation of sensory effects. While elegant, the "strong" version of this theory—proposing direct action-effect associations—is often criticized for being too simplistic for real-world complexity. This study evaluates a computational model of Strong IMT across tasks of varying difficulty. Results indicate that while IMT is a powerful mechanism for bootstrapping simple actions, it requires additional processes like planning and consolidation to handle complex sequences and temporal delays.
Introduction
Ideomotor Theory (IMT) posits a bidirectional link between actions and their sensory consequences. Through experience, we learn that a specific motor command (action) leads to a specific outcome (effect). Later, simply imagining the outcome can trigger the motor command. This "Strong IMT" hypothesis suggests a direct, unmediated trigger. However, critics argue that sensorimotor noise, redundancy, and complex sequences make such a direct mechanism computationally insufficient. This paper addresses these concerns by testing a minimalist neural network model of IMT in simulated environments.
Methods: The Computational Model
The authors implemented a "minimalist" model consisting of a single-layered neural network. The architecture connects Action nodes to Effect nodes via unsupervised, Hebbian-like associative learning. When an action is performed and an effect is perceived, the connection is strengthened. During goal-directed execution, activating an Effect node (the goal) spreads activation back to Action nodes, where a "winner-takes-all" mechanism selects the strongest association.
The model was tested in a "Throwing Task," where a simulated agent learns to throw a ball to specific heights. Complexity was varied by introducing noise, redundant actions (multiple ways to reach a goal), and sequential requirements.
Key Results and Scaling
The model demonstrated remarkable efficiency in "bootstrapping" simple tasks. In basic one-to-one mappings, it reached near 100% success rapidly. However, performance was sensitive to the scale of the action-effect space and the presence of noise.
As shown in Figure 5A, increasing the number of potential actions and effects significantly slows the learning rate, though the model eventually reaches high proficiency. Figure 5B illustrates that while the model is robust to individual noise sources (like noisy representations or actions), the combination of all noise sources creates a substantial "learning tax," requiring more episodes to achieve mastery.
Limitations and the Need for Planning
While Strong IMT excels at immediate, simple mappings, it faces a "temporal wall." If an effect is delayed, the association fails unless a "trace decay" mechanism keeps the action node active. Furthermore, the model struggled with long sequences; in a 12-node task, pure IMT could not reliably reach goals requiring more than 2-3 steps.
To overcome these hurdles, the authors extended the model with a Planning mechanism. By chaining learned associations "offline," the success rate for complex sequences improved from mediocre to nearly 90% within 200 trials, a massive improvement over the 10,000 trials required without planning.
Conclusion
Computational modeling suggests that Ideomotor Theory provides a robust framework for the initial stages of action-effect learning. However, "Strong IMT" is likely only one component of a larger cognitive architecture. For complex, real-world behaviors, the direct associations formed via IMT serve as the essential building blocks for higher-level planning and consolidation processes.
Summary updated