A Framework for Resolving Motivational Conflict via Attractor Dynamics
193
An influential conceptual model of motivation was introduced by Lorenz
[10]. Accordingly, independent reservoirs are filled by ‘energies’ that are actionspecific, and when a corresponding energy threshold is exceeded a valve opens to
release the corresponding behavioral pattern. Theoretical investigations of motivational conflict often extend this idea to additionally consider direct interactions
between motivations via cross-inhibition [12]. Most early investigations incorporated analogies from control theory [18], and in particular used feedback loops
to implement Lorenzian energy build-up with competition between motivations
driven by internal deficits. Deficits usually inhibit one another directly [11] or
bias a decision switch based on ‘tendencies’ derived from internal state [7]. Such
models can be shown to reproduce a wide range of motivational phenomena, but
they have been difficult to map directly onto neural systems.
Here we develop a model of how multiple conflicting motivations, each formulated in homeostatic terms, may be resolved to generate appropriate animal (or
robot) behaviour. The model is considered first in theoretical terms, and then
in terms of simulated robot behaviour.
2 Behaviour Under Motivational Conflict
The proposed model has four stages: Internal physiological state, motivational
dynamical system, behavioural selection and pattern expression.
Internal Physiological State
We describe the internal physiological state of the agent as a dynamical system.
What are referred to in the literature as deficits, are encoded in the state vector
x ∈ R
n . For some motivational systems, e.g., thirst or thermoregulation, the state
vector can be associated with a physical quantity such as the amount of water or
body heat. For others, such as aggression, it can be related to the accumulation
of an action-specific ‘energy’, in the Lorenzian sense [1]. The internal state vector
evolves according to the following dynamical law,
˙
x(t) = −f (x(t)) + g(u(t)) + h(a(t)).
(1)
The first term on the right, −f (x), represents the decay of the energy, corresponding to the Lorenzian model, with the function f specifying the nature
of the decay (e.g., zero-order, first-order etc.) as well as potential interactions
between homeostatic systems. The second term represents an external input
u(t), with the function g allowing for a linear or nonlinear transformation of
that input (i.e., to represent absorption or thermal conductivity dynamics etc.).
The final term represents an autonomic homeostatic process, a(t), with h similarly enabling linear or nonlinear transformations.
The physiological processes we consider are assumed to evolve on slower
timescales than that which characterises the behavioural responses, in a close
submanifold of R
n , given the existence of physiological limits for all processes,
i.e., concentrations can not be negative. To illustrate, consider the following
examples.
193
An influential conceptual model of motivation was introduced by Lorenz
[10]. Accordingly, independent reservoirs are filled by ‘energies’ that are actionspecific, and when a corresponding energy threshold is exceeded a valve opens to
release the corresponding behavioral pattern. Theoretical investigations of motivational conflict often extend this idea to additionally consider direct interactions
between motivations via cross-inhibition [12]. Most early investigations incorporated analogies from control theory [18], and in particular used feedback loops
to implement Lorenzian energy build-up with competition between motivations
driven by internal deficits. Deficits usually inhibit one another directly [11] or
bias a decision switch based on ‘tendencies’ derived from internal state [7]. Such
models can be shown to reproduce a wide range of motivational phenomena, but
they have been difficult to map directly onto neural systems.
Here we develop a model of how multiple conflicting motivations, each formulated in homeostatic terms, may be resolved to generate appropriate animal (or
robot) behaviour. The model is considered first in theoretical terms, and then
in terms of simulated robot behaviour.
2 Behaviour Under Motivational Conflict
The proposed model has four stages: Internal physiological state, motivational
dynamical system, behavioural selection and pattern expression.
Internal Physiological State
We describe the internal physiological state of the agent as a dynamical system.
What are referred to in the literature as deficits, are encoded in the state vector
x ∈ R
n . For some motivational systems, e.g., thirst or thermoregulation, the state
vector can be associated with a physical quantity such as the amount of water or
body heat. For others, such as aggression, it can be related to the accumulation
of an action-specific ‘energy’, in the Lorenzian sense [1]. The internal state vector
evolves according to the following dynamical law,
˙
x(t) = −f (x(t)) + g(u(t)) + h(a(t)).
(1)
The first term on the right, −f (x), represents the decay of the energy, corresponding to the Lorenzian model, with the function f specifying the nature
of the decay (e.g., zero-order, first-order etc.) as well as potential interactions
between homeostatic systems. The second term represents an external input
u(t), with the function g allowing for a linear or nonlinear transformation of
that input (i.e., to represent absorption or thermal conductivity dynamics etc.).
The final term represents an autonomic homeostatic process, a(t), with h similarly enabling linear or nonlinear transformations.
The physiological processes we consider are assumed to evolve on slower
timescales than that which characterises the behavioural responses, in a close
submanifold of R
n , given the existence of physiological limits for all processes,
i.e., concentrations can not be negative. To illustrate, consider the following
examples.