Материал: искусственный интеллект

Hamiltonian-QEM 67

3 The Hamiltonian Based QEM Model

In essence the Hamiltonian based QEM model describes in two subsequent temporal stages how the belief state of the participants evolves through the experimental paradigm. This change of the belief state is described by two distinct Schr¨odinger evolutions. In the first stage the participant is presented with a cue originating from one of the four lists - three of them with targets, one with distractors. First, the participant processes this incoming information to change her initial ‘uniform’ belief state into a state informed by the presented cue and her memory. Subsequently this state of belief is then further evolved due to the processing of the information of the probe. The information of the probe allows for a response bias or description-dependence. We expect the latter evolution to be an attenuation of the first stage recognition phase.

State Vectors. In line with Brainerd, Wang and Reyna’s development of QEM, the state vector is expressed on the orthogonal basis (V1, V2, V3, G, N ). The model thus provides a dedicated dimension for verbatim support for each list, and a dimension for gist support shared for all lists. The last dimension is dedicated to support for non-related items or distractors. In our dynamical development of QEM each state vector is modulated according the cue and probe combination to which the subject is exposed

Ψprobe|cue(t) = [ψp|cV1 (t), ψp|cV2 (t), ψp|cV3 (t), ψp|cG(t), ψp|cN (t)]τ ,

amounting to sixteen distinct states in the present experimental paradigm.

The Initial State Vector. In the basis (V1, V2, V3, G, N ), the generalized initial state vector is expressed as

where we restrict the parameter g [−1, 1]. This initial belief state of the subject reflects to certain extent the fact - of which the subject is informed - that in this experiment half of the cues are non-studied, and half of them come from the three studied lists (p. 419, [7]). This form also implements the idea that the subject at the start has a latent tendency for acceptance of the cues. This form of the initial state expresses that a cue from a studied list elicits on average (1 − g2)/6 acceptance probability from verbatim trace and g2/2 acceptance probability from the gist trace.2 The parameter g therefor indicates the preponderance of gist in

2Without taking into account of dynamics for the e ect of cue or probe, but still applying the measurement projections Eq. (6), the amount of gist g in the initial state shows a latent tendency for overdistribution

p0(L1) + p0(L2) + p0(L3) + p0(N ) = 1 + g2

with p0(N ) = 1/2. Clearly the experimental description ‘half of the cues are N , the other half originate from the lists’, cannot be implemented exactely due to overdistribution.

68 J. B. Broekaert and J. R. Busemeyer

the initial state for a given cue. It should be noticed however that the initial state itself is not measured on. In the Hamiltonian model a dynamical evolution transforms the initial belief state till the point of measurement. Our present assessment of the initial component amplitudes therefore only concern a latent tendency. We will at present fix the initial state to correspond mathematically to equal weighting of acceptance support by verbatim trace and by gist trace for the three studied lists. The initial state then reads, with g = 0.5 in Eq. (5),

Measurement Projectors. The subject’s response to the probes Li?, or L123? for a given cue of Lj , are obtained by applying the measurement operators on the

final state Ψprobe|cue. The measurement operators are projector matrices which select the components of the final outcome vectors for the specific response.

The projector for e.g. L1? must select both the dedicated verbatim amplitude ψL1|cV1 and the gist component ψL1|cG. In (V1, V2, V3, G, N ) ordered Hilbert space the corresponding projector ML1? will thus collapse the outcome state to a subspace spanned on the (V1, G) basis, and is implemented by a matrix with diagonal elements (1, 0, 0, 1, 0) and 0 otherwise. All projectors for the three list paradigm are implemented accordingly:

We notice the acceptance projectors commute, but are not orthogonal as they all include the gist component in their projective subspace;

Trueblood and Hemmer [25], and Denolf and Lambert-Mogliansky [16] point out QEM’s orthogonality of verbatim, gist and non-related features need not necessarily be retained (cfr model specifications in Sect. 1). In our dynamical extension of QEM a query probe engenders its proper dynamics, therefor repeated application of projectors, without intermediate evolution, will not occur.

Acceptance Probabilities. In quantum-like models probabilities are given by the squared length of the projected outcome vectors. Using the projector operators, Eq. (6), the resulting acceptance probabilities for a given probe Li? after a given cue Lj , are explicitly given by the expressions:

Hamiltonian-QEM 69

Notice that the probe index i runs from 1 to 3, while the cue index j runs from 1 to 4 since it includes the non-studied list L4.

With the explicit expressions of the acceptance probabilities we obtain, using some algebra, the potential for super-additivity in the Hamiltonian-QEM model by means of the unpacking factor, for a given cue c;

From which follows that - given the positive gist balance in the second fraction - most often the Hamiltonian-QEM model will predict subadditivity in the three list paradigm. However, like in the Hamiltonian QEM model for the item false memory paradigm of Broekaert and Busemeyer [11], we find the model can in principle also account for cases that satisfy the additivity of disjoint sub-events, or violate it in super-additive manner.

Hamiltonians. The dynamical evolution of the state vectors is determined by the Hamiltonian operators. The Hamiltonian reflects the cognitive processing which is engendered by the information in the word cue - di erently for a cue from the studied lists or the unrelated list. In the present three-list paradigm these operators are constructed along four transports - or re-allocations - between components of the belief state; (i) between gist-based component and non-cue verbatim-based component (G ↔ V¬i), (ii) between gist-based component and non-related component (G ↔ N ), (iii) between cue verbatim-based component and non-cue verbatim-based component (Vi ↔ V¬i), and (iv) between cue verbatim-based component and non-related component (Vi ↔ N ). The Hamiltonians are constructed by combining the o -diagonal parametrised Hadamard gates of each transport [11]. The Hamiltonian parameter γ controls the transport of acceptance probability amplitude from non-cue verbatim-based components towards the gist-based component, or back. Similarly γ regulates transport of acceptance probability amplitude of the non-related component from or towards the gist-based component. Precisely the strength γ of this dynamic will be made use of to adjust for the four distinct types of word cues ‘high frequency & concrete’ (HFC), ‘high frequency & abstract’ (HFA), ‘low frequency & concrete’ (LFC), ‘low frequency & abstract’ (LFA). Brainerd and Reyna suggest abstract words have weaker verbatim traces than concrete words and lowfrequency words have weaker verbatim traces than high-frequency words [5, 7, 8]. A tendency which suggests gist based transport from the non-related component

to vary—namely {γHF C , γHF A, γLF C , γLF A}—for these distinct types of word cues.

The parameter ν controls transport of acceptance probability amplitude between non-cue verbatim-based components and the cue verbatim-based component. Similarly ν controls the transport of the acceptance probability ampli-

70 J. B. Broekaert and J. R. Busemeyer

tude of the non-related component from or towards the cue verbatim-based component.

In our present development we demonstrate the parameters fulfil their intended transport: the parameters γ and γ for gist-based transport, and ν and ν for cue verbatim-based transport. We could have implemented an e ect of list order by distinguishing for each list the transports of the non-related com-

ponent (γi,Abs.type, νi ) for forgetfulness variation per list, or by distinguishing the transports of the target verbatim-based component (γi, νi) for gist-verbatim

diversified acceptance variation per list.

Diminishing the strength of transports, by quenching the driving parameters, will let outcome acceptance beliefs tend to concur with the acceptance beliefs inherent to the initial state.

First Temporal Stage. The presentation of the word cue starts the memory process of recollection and familiarity in the subject. For cues from studied lists the hamiltonians H1c, H2c and H3c engage verbatim-based and gist-based belief. For non-studied cues from list 4, the hamiltonian H4c engages for the non-related belief.

Second Temporal Stage. The presentation of the probe starts the subsequent decision process in the subject. The outcome of the cue processing and the subsequent probe can be confirmatory or dissonant. When probing for a studied list, the Hamiltonian Hip for a probe Li? is equated to the Hamiltonian Hic for processing a cue from Li, in which the driving parameters are now attenuated by a factor κ. When probing for the union of lists, the dedicated Hamiltonian H123p for probe L123? is equated to the sum of the separate Hamiltonians Hip for processing of studied cues from Li, in which the driving parameters are again attenuated by the factor κ. This separable construction of the Hamiltonian reflects the cognitive processing of the union list consists of the parallel processing of membership to the three separate lists, and results in the summed attenuated transport between the non-related component and equally all verbatim components and the gist component.

Parameters. In recapitulation; the Hamiltonian QEM model uses 5 parameters {μ, μ , γ, γ , κ} to describe the dynamics of the subject’s belief state to evolve from her prior partially informed expectation to the final decision of acceptance of the cue, accounting for 16 acceptance probabilities p(Li?|Lj ). In Sect. (4), we fitted the parameters for the four distinct types of frequency and concreteness of the cue words HFC, HFA, LFC and LFA. Between these sets we only changed the gist transport parameter γ (G ↔ N ) in the Hamiltonians, respectively γHFA, γHFA, γLFC and γLFA, and otherwise maintained the same Hamiltonian drivers γ, ν, ν and κ. In total therefor 8 parameters are needed to predict the 64 values of the acceptance probabilities.

Unitary Time Evolution. In quantum-like models the Hamiltonian is the operator for infinitesimal time change of the belief state. The operator U (t) for time propagation over a time range t is given by

The fully evolved belief state - a solution of the Schr¨odinger equation - is obtained by applying the unitary time operator U (t) to the initial state. An inherent feature with Hamiltonian quantum models is the appearance of oscillations of probability over time. In quantum mechanics, finite dimensional and energetically closed systems are always periodical. Therefor the initial belief state will re-occur after the proper time period of the system. In previous work we have argued a third temporal stage, closing the experimental paradigm, should implicitly be supposed in which all driving parameters are set equal to zero [11]. The time of measurement t is taken equal to π2 for each stage [14].3 The final outcome state vector is thus obtained by concatenating both propagators on the initial state, Eq. (5),

The processing of cue c in the first stage occurs till t = π2 , the processing of probe p? in the second stage takes another time range of π2 . The time evolution of the acceptance probabilities from initial state processing of cue and processing of probe is shown in Fig. 2.

3 The Hamiltonian parameters are dependent on the choice of measurement time.