Artificial Emotions: Good Bye Mr. Spock!

Rodrigo Ventura, Luís Custódio and Carlos Pinto-Ferreira ask whether mechanisms inspired by emotion can help an agent make useful decisions under time pressure. Their proposal separates a rich representation of a situation from a simpler assessment of whether it is desirable. A small implementation illustrates how those assessments can be learned and revised; it does not test a robot against an unemotional counterpart.^[1]

Emotion as a decision mechanism

The authors distinguish two research aims. A behavioral approach adds recognizable emotional expression to improve interaction. Their functional approach instead asks what emotion-like mechanisms might contribute inside a decision system. A system need not imitate detailed human physiology to investigate that question. Their motivation is that an exhaustive reasoning process may deliver a correct answer too late to be useful; the paper proposes a computational role for quick assessment rather than demonstrating that emotions are necessary for every form of intelligence (§§1–2).^[1]

Two representations, two connections

In the proposed architecture, an external stimulus reaches two processors at once. The cognitive processor produces a generalized image: a rich, structured representation, potentially visual or auditory. The perceptual processor produces a vector of desirability: a simpler set of built-in assessments such as positive versus negative or desirable versus avoidable. The vector sacrifices descriptive richness to be easier to handle.^[1]

Two connections are intended to make the representations useful together. Marking associates the rich image with its desirability assessment, inspired by Damasio’s somatic-marker hypothesis. Indexing would allow a primitive assessment to retrieve a relevant cognitive context quickly. The authors’ approaching-object example separates an immediate cue to move away from the later work of identifying what is coming. It illustrates the design rationale, rather than a timed experiment with that architecture (Figure 1; §2).^[1]

What was implemented

The simplified system uses two inputs, called vision and environment, and implements marking only. A new visual image recalls similar stored images into working memory. Each recalled image has a learned marker and a relevance value reflecting similarity. The system combines that marker with the current environmental input; relevance determines how much weight the old marker receives. It then stores the updated image and marker back in main memory (Figure 2; §3).^[1]

This produces both an assessment called a “feeling” and a positive, negative or neutral response. The paper explicitly uses that vocabulary because of the design’s inspiration, without claiming that humans feel in the same way. The diagram’s “body” is a placeholder for the computed somatic response, rather than evidence of a simulated biological body.^[1]

The reported demonstrations present image–environment pairs. Later, a similar image presented without environmental input recalls its learned marker with high relevance. Repeatedly presenting the image alone makes the association fade; changing the environmental input changes the marker. These qualitative examples illustrate recall and adaptation. They supply no reported sample-size comparison, error-rate table or timing benchmark establishing an advantage over another decision architecture (§3).^[1]

The contribution and its open test

The institute’s publication catalogue places this exact eight-page manuscript in the IBERAMIA 1998 proceedings, pp. 395–402. Its website’s 2015 upload directory is not the original publication year.^[1]

For Machine Intelligence, the distinctive proposal is to bind rich representations to inexpensive, revisable assessments. The demonstrated learning mechanism is narrower than the full architecture: indexing remains unimplemented, and successful association does not establish competent action across complex environments. The paper offers a design direction for emotion-inspired decision-making, leaving comparative usefulness, appropriate learned assessments and the consequences of acting on them to further evaluation.^[1]