The Cognitive architecture of the AHA agents

The AHA Model: Adapted Heuristics and Architecture: This is a large scale simulation model developed at the Theoretical Ecology Group, University of Bergen, that implements a general decision-making architecture in evolutionary agents.

The cognitive architecture of the agent is represented on the scheme below. It includes several functional units, each representing specific motivation or emotion state: survival circuit.

Note

Note that the scheme below includes three survival circuits for the states A, B, C. For simplicity, there are also only three stimuli S1, S2, S3.

The agent perceives the outer and its own inner environment, obtaining perception (signal) values. The agent also has several motivation (emotional) states, such that only one can be active at any time step of the model. Thus, the states compete for this position. The winning motivation (emotion) becomes the dominant emotional state of the agent, its Global Organismic State. This "Global Organismic States" of the agent determines how the agent weights different options during its decision making process and therefore determines what kind of behaviour (action) it will execute.

Perception and appraisal

The agent obtains "perceptions" (P) from its external and internal environments. These perceptions are fed into the "Appraisal" modules, separate for each of the motivation/emotion state.

Here perception signals are first fed into the "neuronal response functions" (based on the sigmoidal function commondata::gamma2gene()). The neuronal response function is a function of both the perception signal (P) and the genome (G) of the agent. Perception signal is also distorted by a random Gaussian error. Each neuronal response function returns the neuronal response (R) value.

The neuronal responses R for each stimulus are summed for the same motivation module to get the primary motivation values M1) for this motivational state. These primary motivations can then be subjected to genetic or developmental (e.g. age-related) modulation, resulting in the final motivation values Mf. Such modulation could strengthen or weaken the motivation values and therefore shift the outcome of competition between the different motivational states. In absence of modulation M1 = Mf.

Global Organismic State

Final motivations Mf for different motivations (emotions) are competing, so that the winning state that is characterised by the highest final motivation value becomes the dominant emotional state of the agent: its "Global Organismic State" at the next time step. Additionally, the final motivation value of this state becomes the GOS arousal level.

The competition mechanism is complex and dynamic. It depends on the current arousal level, such that relatively minor fluctuations in the stimuli and their associated motivation values are ignored and do not result in switching of the GOS to a different state.

Furthermore, the relative difference (surplus) that the competing motivation must have to win competition against the current state depends on the current level of GOS arousal. If the current arousal level of the agent is relatively low (motivation or emotion is weak), a competing state must exceed a relatively higher threshold to win. However, if the current arousal level is very high (high motivation or emotion), a competing state can win even if it only slightly exceeds the current arousal level. Thus, the emotional state of the agent is characterised by a degree of continuity or "inertia" and such inetria is lower the higher is the current level of arousal.

Attention

Whenever the agent has a specific Global Organismic State, this state also affects the agent's perception. All the perception inputs that belong to motivations other than the currently dominant (i.e. the current GOS) are suppressed by attention weights. For example, if the motivation B is the GOS, all the perception values linked with Motivation A and C are suppressed. The suppression weights are proportional to the current GOS arousal of the agent.

Thus, the attention mechanism effectively filters out or "focus" the agent on the stimuli that are linked with the current dominant emotional state. Moreover, the stronger is the current GOS arousal, the stronger is such attention focusing.

The Perception-to-Arousal path

This process, perception → neuronal response → motivation → GOS → arousal is repeated at each time step of the model as the agent acts in (e.g. moves through) its stochastic environment. In effect, the dominant motivational and emotional state of the agent changes adapting to to the latest changes in the inner and external environment.

Self-predictive decision making

Furthermore, the same processes (and computer code procedures) are also evoked when the agent is making the decision about what behavioural action to choose at each time step.

Basically, the agent predicts what would be its perceptions and, for each of the behavioural action available, runs the same process (perception → neuronal response → motivation → GOS → arousal) and finally selects the behaviour that would result in the lowest predicted GOS arousal. Perceptions in this process are predicted from the agent's internal or local external environment ("fake" perceptions). They are also subjected to attention suppression, however, attention weights are transferred from the current Global Organismic State.

Thus, decision making of the agent is based on predicting one's own emotional state. The emotional arousal becomes a common currency in decision making.

Goal directed behaviour

The cognitive architecture implemented in the AHA model effectively produces goal-directed behaviour in the agents. The "goal" is defined by the evolutionary process, the Genetic Algorithm. Specifically, the target "goal" is to survive, grow, accumulate energy and reproduce. The agents that do best in this respect pass their genes into the next generation.