202
A. Jimenez-Rodriguez et al.
A theoretical analysis revealed how the motivational state of the agent can
depend on its internal state, allowing us to recast some motivational phenomena
described in the animal behaviour literature. For example, we showed how the
fixed points of the motivational dynamical system can disappear, appear, or
merge, as a function of internal state. We showed how merging of the fixed
points can generate the kinds of ambiguous motivational states that have been
observed in some animals [1]. And by incorporating stochasticity, we showed how
the model can account for phenomena such as time-sharing [7,14] in terms of
spontaneous transitions and disinhibition [15].
Our model resembles several classical approaches [19] to this problem. For
example, the changes in the underlying physiological dynamics that are described
correspond to Lorenzian energy accumulation, but with the valve that releases
behaviour recast as a bistability in the motivational state space. However, early
models of motivational conflict typically involve positive feedback loops according to a control theory framework [7,17,18], and our model does not include
explicit feedback loops, with the control instead directly implemented through
the effect of the environment on the internal state of the agent.
The stochastic motivational system resembles accumulator models [12].
Indeed, given two stochastic motivations, v 1 and v 2 , as in [12], the motivational
particle, ρ, could be associated with the difference v 1 − v 2 of such accumulators.
However, interaction between motivations is assumed not to happen explicitly
in our model (i.e. no explicit cross-inhibition), but it is instead assumed to be
mediated via an external field.
Our decision to embed motivational attractors in a metric space corresponds
to an assumption that the underlying space in which behaviours reside has a
definite topology, i.e., that behaviours can be meaningfully ordered and that the
ordering determines the interactions between behaviours that may be observed
in various conflict scenarios. As such, it should be possible to devise experiments
to determine a definite pattern of displacement phenomena for a given animal
(or species).
Finally, we note that the interaction between the motivational state and the
readouts resembles observations of the interactions between the lateral hypothalamus and the ventral tegmental area of the mammalian brain, and that the
tendency readout could be potentially associated with the dopamine signals that
relate to value in these areas [4]. Further connections with the neurobiology of
decision-making in animal brains, with a focus on possible relationships with
reinforcers and reward, will be explored in future work.
Simulation code is available at https://github.com/ABRG-Models/
MammalBot/.
Acknowledgments. This work was supported by the EU H2020 Programme as part
of the Human Brain Project (HBP-SGA2, 785907).
A. Jimenez-Rodriguez et al.
A theoretical analysis revealed how the motivational state of the agent can
depend on its internal state, allowing us to recast some motivational phenomena
described in the animal behaviour literature. For example, we showed how the
fixed points of the motivational dynamical system can disappear, appear, or
merge, as a function of internal state. We showed how merging of the fixed
points can generate the kinds of ambiguous motivational states that have been
observed in some animals [1]. And by incorporating stochasticity, we showed how
the model can account for phenomena such as time-sharing [7,14] in terms of
spontaneous transitions and disinhibition [15].
Our model resembles several classical approaches [19] to this problem. For
example, the changes in the underlying physiological dynamics that are described
correspond to Lorenzian energy accumulation, but with the valve that releases
behaviour recast as a bistability in the motivational state space. However, early
models of motivational conflict typically involve positive feedback loops according to a control theory framework [7,17,18], and our model does not include
explicit feedback loops, with the control instead directly implemented through
the effect of the environment on the internal state of the agent.
The stochastic motivational system resembles accumulator models [12].
Indeed, given two stochastic motivations, v 1 and v 2 , as in [12], the motivational
particle, ρ, could be associated with the difference v 1 − v 2 of such accumulators.
However, interaction between motivations is assumed not to happen explicitly
in our model (i.e. no explicit cross-inhibition), but it is instead assumed to be
mediated via an external field.
Our decision to embed motivational attractors in a metric space corresponds
to an assumption that the underlying space in which behaviours reside has a
definite topology, i.e., that behaviours can be meaningfully ordered and that the
ordering determines the interactions between behaviours that may be observed
in various conflict scenarios. As such, it should be possible to devise experiments
to determine a definite pattern of displacement phenomena for a given animal
(or species).
Finally, we note that the interaction between the motivational state and the
readouts resembles observations of the interactions between the lateral hypothalamus and the ventral tegmental area of the mammalian brain, and that the
tendency readout could be potentially associated with the dopamine signals that
relate to value in these areas [4]. Further connections with the neurobiology of
decision-making in animal brains, with a focus on possible relationships with
reinforcers and reward, will be explored in future work.
Simulation code is available at https://github.com/ABRG-Models/
MammalBot/.
Acknowledgments. This work was supported by the EU H2020 Programme as part
of the Human Brain Project (HBP-SGA2, 785907).