bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 5쪽 중 3쪽
dded to avoid division by zero. The parameter p associated with the potential function controls the
spread of the rewards. The fiducial parameter values are p = 3,c = 0.75,ϵ = 0.01, and ∆N = 10 for the time window
over which the standard deviationσ(T ) is evaluated.
As discussed above, reputers require scoring according to their consensus in reporting the ground truth. The naive
way of doing this is to calculate the stake-weighted average of all losses reported by a reputer, and to add the rewards to
their stakes. However, this creates a runaway effect towards increased centralization, where the reputer with the highest
stake has the largest weight in setting the consensus, thereby receiving the highest rewards and further increasing their
stake advantage. This can be remedied by using an adjusted stake to set the weight of each reputer when calculating the
consensus, where the weight saturates above a certain fraction of the stake. Specifically, Allora assigns an adjusted stake
for calculating the consensus as
ˆSim = min
( NraimSim∑
maimSim
, 1
)
, (30)
whereNr is the number of reputers and aim is a listening coefficient defined below, which falls in the range [0, 1] and
represents a weight that the network associates with each individual reputer depending on its historical performance. This
function returns unity for above-average stake fractions ( aimSim/∑maimSim > 1/Nr), whereas for below-average
stake fractions (Sim/∑mSim < 1/Nr) it increases linearly from zero at Sim = 0 to unity at aimSim/∑maimSim =
1/Nr. This formulation ensures that the consensus calculation is not susceptible to a majority attack, as reputers with
above-average stake have equal weight in setting the consensus. The magnitude of their stake influences the rewards
ADI 1, 1–19 (2024)
10 Kruijssen et al.
received, but cannot be used to further increase their influence in defining the reference point for the reward distribution.
This avoids the runaway effect leading to ever-increasing centralization.
With this adjusted definition, we collect all losses reported by a reputer as
Lim ={Lim,Lijm,Likm,L−im,L−lim,L+kim}, (31)
and define the consensus loss vector as
log Li =
∑
m ˆSim log Lim∑
m ˆSim
. (32)
The reward received by each reputer is set by a combination of its stake and the Euclidean proximity of its reported losses
to the consensus loss vector. We score the proximity to consensus as
Tim =
[‖log(Lim/Li)‖
‖log Li‖ +ϵr
]−1
, (33)
where the first term expresses the relative proximity andϵr = 0.01 is a small tolerance quantity used to cap reputer scores
at infinitesimally close proximities. With these definitions, we can now calculate the listening parameters, which we
obtain by gradient descent. As the objective function, we use the stake-weighted total consensus score:
Ti =
∑
mSimTim∑
mSim
. (34)
The listening coefficients are initialized at unity and are updated following an iterative process:
aim→aim +λd lnTidaim , (35)
where λ is the learning rate and d lnTi/daim is the relative gradient. The iterative update of Equation 35 is carried
out each epoch until the relative gradient reaches d lnTi/daim < 0.001, or until the maximum of 1/λ iterations is
reached. Whenever the update of the listening coefficients of Equation 35 decreases the fraction of stake that is listened
to below∑maimSim/∑mSim < 0.5, the differential is instead interpolated to ensure ∑maimSim/∑mSim = 0.5.
The resulting listening coefficients carry over into the next epoch, where they are updated using the same process. This
gradient descent mechanism enforces consensus and ensures that the network is robust against minority attacks, where
reputers incorrectly report on the losses of favored workers. The network learns the listening coefficients aim to ensure
that such dishonest reputers are quickly silenced.
For a total reward allocated to reputers per time step ( Wi, specified in §4.2), we now calculate the fraction of the
rewards allocated to reputers that is received by each individual reputer as
wim = (SimTim)p∑
m (SimTim)p, (36)
where the multiplication by stake and consensus score ensures the appropriate dependence of the reward fraction on both
quantities, and the parameter p grants the ability to modify the reward spread (we adopt a fiducial value of p = 1). Each
reputer’s reward is added to their stake, so that the stake is constituted by a combination of monetary commitment and
historical performance:
Si+1,m =Sim +wimWi, (37)
whereWi is the total reward allocated to reputers (see §4.2). This way, poorly performing reputers experience dilution
of their stake and weight, whereas accurate reputers can grow their influence. Another major advantage of this form of
reward payments to reputers is that the topic is secured by capital in rough proportion to its value.
4.2 Reward Division between Network Tasks
The system described above governs the distribution of rewards among individual participants performing the same task.
Next, we define how the total rewards emitted to a topic t in a given time step Et,i are divided among the three classes
of tasks within the topic. The differentiation in incentive structures between the classes requires the definition of a new
objective function that is relevant in each of the three cases. The common objective of the network is to incentivize
decentralization, and this can be quantified for each class of tasks by considering the entropy of the reward distribution
among participants performing that task. The entropy increases for larger numbers of participants and for more equal
reward distributions.
We define a modified entropy for each class ({Fi,Gi,Hi} for the inference, forecasting, and reputer tasks, respec-
tively) to quantify its degree of decentralization:
Fi =−
∑
j
fij ln (fij)
(Ni,eff
Ni
)β
, Gi =−
∑
k
fik ln (fik)
(Nf,eff
Nf
)β
, Hi =−
∑
m
fim ln (fim)
(Nr,eff
Nr
)β
, (38)
ADI 1, 1–19 (2024)
Allora: a Self-Improving, Decentralized Machine Intelligence Network 11
where we have defined modified reward fractions per class as
fij = ˜uij∑
j ˜uij
, fik = ˜vik∑
k ˜vik
, fim = ˜wim∑
m ˜wim
. (39)
Here, the tilde over the reward fractions indicates that we have applied an exponential moving average:
˜uij =αuij + (1−α)˜ui−1,j, ˜vik =αvik + (1−α)˜vi−1,k, ˜wim =αwim + (1−α) ˜wi−1,m, (40)
to enable the entropy to include some record of the recent historical degree of decentralization rather than just the last
time step. As before, we adopt a fiducial value ofα = 0.1.
Finally, the entropy requires modification by the number ratio term in Equation 38, because the increase of the actual
entropy with lnN might otherwise incentivize a sybil attack, in which any class of topic participants add many copies
of themselves to the network in an attempt to boost their entropy and class reward allocation. The number ratio term in
Equation 38 has been added to prevent this. It contains an effective number of participants, which is defined as
Ni,eff = 1∑
jf2ij, Nf,eff = 1∑kf2ik, Nr,eff = 1∑mf2im. (41)
For a uniform reward distribution, Neff = N by definition. For strongly unequal reward distributions, Neff≪ N. With
the addition of the (Neff/N)β term in Equation 38, any sybil attack would necessarily dilute the individual rewards of the
sybil, which would increase the number of participants while keeping their effective number approximately constant. As
a result, the number ratio term (Neff/N)β would decrease and help maintain constant entropy. Tests of this formulation
show thatβ = 0.25 achieves this nearly exactly, and we adopt this as a fiducial value.
Given a total reward emitted to a topict within a time stepEt,i, we use the entropies to define the part of the rewards
that is allocated to each class as
Ui = (1−χ)γFiEt,iFi +Gi +Hi , Vi = χγGiEt,iFi +Gi +Hi, Wi = HiEt,iFi +Gi +Hi. (42)
In other words, each class of activity within a topic is allocated a reward proportional to the fraction of the total modified
entropy (Fi +Gi +Hi) that is generated by that class. The rewards for the inference and forecasting tasks each contain
an additional factor ( (1− χ)γ and χγ, respectively). These factors are included, because the reward split between
the inference and forecasting tasks should additionally acknowledge the added value provided by the forecasting task.
Fundamentally, the inference task is the engine of the Allora network; without any inferences to start with, there would
be no network inference. By contrast, the network can function without the forecasting task, but is included to create
context awareness and increase the accuracy of the network inference. Therefore, it is reasonable to modulate the reward
allocation between both worker tasks according to the relative utility of the forecasting task.
The forecasting task utility χ is measured using a definition analogous to the one-out performance scores in Equa-
tion 22. We subtract the logarithm of the loss of the complete network inference (Li) from that of the naive network (L−i ),
which is obtained by omitting all forecast-implied inferences:
Ti = logL−i − logLi. (43)
The performance score of the entire forecasting task Ti is positive if the removal of the forecasting task would increase
the network loss, and is negative if its removal would decrease the network loss. We then define a modified ratio between
the forecasting task performance score and the sum of all inference scores as
τi≡αTi− min (0, maxjTij)|maxjTij| + (1−α)τi−1, (44)
withα = 0.1, and apply a piecewise linear transformation to set the forecasting task utility within the rangeχ = (0.1, 0.5):
χ =
0.1 ifτi < 0,
0.4τi + 0.1 if 0≤τi < 1,
0.5 ifτi≥ 1.
(45)
This represents a linear transition from χ = 0.1 toχ = 0.5 in the rangeτi = [0, 1]. We can now multiply the allocation
of class rewards for the inference and forecasting tasks by (1−χ) andχ, respectively, and then renormalize with a factor
γ to ensure that the total reward allocated to workers (Ui +Vi) remains constant (otherwise, this would go at the expense
of reputers). It is straightforward to demonstrate that this normalization factor reads
γ = Fi +Gi
(1−χ)Fi +χGi
. (46)
ADI 1, 1–19 (2024)
12 Kruijssen et al.
103
104
Time step rewards [ALLO]
T otal rewards
103
104
Inference task rewards
103
104
Forecasting task rewards
103
104
Reputer rewards
2021-012021-042021-072021-102022-012022-042022-072022-102023-012023-042023-072023-10
Time
103
104
105
106
107
Cumulative rewards [ALLO]
T otal
Inference task
Forecasting task
Reputer task
2021-012021-042021-072021-102022-012022-042022-072022-102023-012023-042023-072023-10
Time
103
104
105
106
107
All workers
Worker 1
Worker 2
Worker 3
2021-012021-042021-072021-102022-012022-042022-072022-102023-012023-042023-072023-10
Time
103
104
105
106
107
All workers
Worker 1
Worker 2
Worker 3
2021-012021-042021-072021-102022-012022-042022-072022-102023-012023-042023-072023-10
Time
103
104
105
106
107
All reputers
Reputer 1
Reputer 2
Reputer 3
Figure 3: Demonstration of the incentive structure of the Allora network, showing how rewards are distributed among
different network participants over time. The left-hand column shows the total rewards given to each of the network task
classes, i.e. the inference task (blue), the forecasting task (cyan), and the reputer task (red), as well as the combined total
in black. The top panel shows these rewards per time step and the bottom panel gives their cumulative sum over time. The
other columns show the same information broken down into rewards for individual participants engaged in the inference
(left-middle), forecasting (right-middle) and reputer (right) tasks.
Substitution of Equation 46 into Equation 42 indeed results inUi +Vi =Et,i(Fi +Gi)/(Fi +Gi +Hi), as desired. The
rewards per class specified in Equation 42 can now be combined with the reward fractions from Equation 27, Equation 28,
and Equation 36 to obtain the absolute rewards paid out to individual network participants:
Uij =uijUi, Vik =vijVi, Wim =wijWi. (47)
Figure 3 illustrates the resulting incentive structure of the Allora network, using the same numerical experiment as
considered in Figure 2 and additionally adopting a stochastic model to describe the total reward emission, which is the
main boundary condition needed to simulate the reward distribution (black line in the left panels of Figure 3). The best-
performing participants (Worker 2 and Reputer 2) receive the largest share of the rewards on average, allowing them to
achieve the highest cumulative rewards at the end of the experiment. Toward the end of the experiment, the spread in
rewards is larger for the reputers than for either of the two worker tasks, indicating that their adjusted entropy is lowest.
As a result, reputers receive the smallest reward share. The allocation of higher cumulative rewards to the inference task
than the forecasting task is mainly due to the low initial utility of the forecasting task, which is visible in Figure 2 as the
steep decrease in the losses of the forecast-implied inferences during the first time steps. The forecasting task needs some
time before it outperforms the raw inferences, which means that initially the inference task receives most of the worker
reward share. This initial difference is partially overcome later on by the small spread in rewards per time step of the
forecasting task and its correspondingly high entropy, which leads to a high reward allocation relative to the other tasks.
The above design represents a complete description of the differentiated incentive structure of a topic within the Allora
network. The described rule set appropriately rewards workers for high-quality inferences obtained from their inference
and forecasting tasks. It also rewards reputers according to their stake and consensus, allowing them to provide economic
security to the topic. Finally, it incentivizes and rewards a high degree of decentralization, where a topic hosts a large
number of participants that all make relevant contributions to network inferences and network security.
5 Allora Tokenomics
5.1 Token Emission Rate
The Allora token (ALLO) is minted by the Allora network to facilitate the exchange of value by network participants.
Specifically, the token holders can choose to engage in the following forms of token utility:
1. The ALLO token can be used to purchase inferences that are generated by the network. Allora adopts a pay-what-
you-want (PWYW) model, where token holders choose the ALLO fee they are paying for an inference (see§5.3 for
details).
2. The ALLO token can be used to pay the access fee for topic creation, or for participating in the network as a worker,
reputer, or network validator. The access fee price is variable.
ADI 1, 1–19 (2024)
Allora: a Self-Improving, Decentralized Machine Intelligence Network 13
3. The ALLO token can be used by reputers and network validators to stake, and by any other token holders to delegate
stake to a reputer or network validator. Staking reputers, staking network validators, and delegating token holders
receive rewards in ALLO.
4. The ALLO token is used by the network to pay out rewards to network participants. For workers, these rewards
are proportional to their unique contribution to the network accuracy. For reputers and network validators, these
rewards are proportional to their stake and the consensus between their output and those of other participants (see
§4). The total amount of ALLO emitted by the network is offset by any fees collected, implying that the ALLO
in circulation can increase (corresponding to inflation) or decrease (corresponding to deflation) depending on the
market dynamics.
The above comprehensively describes the utility of the ALLO token. ALLO tokens will not in any way represent any
shareholding, participation, right, title, or interest in any company, enterprise or undertaking, nor will the tokens entitle
token holders to any promise of dividends, revenue, capital, profits or investment returns.
Token emissions serve the purpose of providing liquidity for reward payouts. The general philosophy adopted by
the Allora tokenomics revolves around two key goals. First, ALLO emissions are subjected to a smoothened form of a
Bitcoin-like schedule (Nakamoto, 2008) to maintain long-term rewards in a limited-supply economy. Second, the annual
percentage yield (APY) earned per staked token should be stable around major token unlocks to disincentivize these
tokens from being dumped on the market. Together, these goals imply a simple emission design that is described as
follows.
The total emission at time stepi is given by
Ei =ei
Nstaked,i
Nepochs
, (48)
whereEi is the total number of tokens emitted as rewards, ei is the monthly emission per unit staked token, Nstaked,i
represents the number of staked tokens, andNepochs is the monthly number of time steps (epochs) during which rewards
are paid out. We define a target monthly emission per unit staked token that the network strives for:
ˆetarget,i = feTtotal,iNstaked,i Ncirc,iNtotal,i, (49)
whereTtotal,i is the total number of tokens held by the network treasury, Ntotal,i = 1, 000, 000, 000 is the total token
supply,Ncirc,i is the circulating supply, andNstaked,i is the staked supply. The factor fe = 0.025 month−1 represents
the fraction of the network treasury that would ideally be emitted per unit time. Equation 49 states that the total emission
per unit time should be equal tofeEtotal,i in the limit where the entire token supply would be in circulation.
If only a small fraction of the circulating supply is staked (i.e. Nstaked,i≪N circ,i), then Equation 49 can result in
an undesirably high APY . Allora therefore limits the APY to 12% and correspondingly limits the monthly percentage
yield (MPY) to 0.95%. This is achieved by limiting the target emission per unit staked token. The MPY depends on the
monthly emission per unit staked token as
ξi =fstakers,iei, (50)
where fstakers,i represents the fraction of the total token emission during the previous epoch that was paid to reward
staking network participants, i.e. reputers and network validators. The maximum monthly emission per unit staked token
given a maximum MPY (ξmax = 0.0095) then follows as
ˆemax,i = ξmaxfstakers,i, (51)
so that the target monthly emission per unit staked token becomes
ˆei = min (ˆetarget,i, ˆemax,i). (52)
While Equation 49-Equation 52 set the target emission per unit staked token, in practice this expression introduces
undesired behavior around major token unlocks. The jumps inNstaked andNcirc may result in APY drops and would thus
incentivize holders to add to the selling pressure normally associated with token unlocks. To decrease the selling pressure
and maintain a healthy economy, Allora adjusts the token emission to stabilize the emission per unit staked token. This
emission smoothing is achieved by applying an exponential moving average to the monthly emission per unit staked token:
ei =αeˆei + (1−αe)ei−1, (53)
whereαe = 0.1 is a parameter setting the degree of smoothing when the update cadence is one month. It can easily be
transformed to other cadences using the conversion:
ˆαe = 1− (1−αe)∆t/month, (54)
ADI 1, 1–19 (2024)
14 Kruijssen et al.
2024-072025-012025-072026-012026-072027-012027-072028-012028-072029-012029-072030-012030-072031-012031-072032-012032-07
Time
100
101
APY [%]
With emission smoothing, with 2% monthly fee revenue
With emission smoothing, no fee revenue
No emission smoothing, no fee revenue
Figure 4: Illustration of the APY earned by staking ALLO holders under the Allora Network tokenomics design. The
figure shows the predicted APY evolution for three different setups, illustrating that the addition of emission smoothing
(from Equation 49-Equation 54) and fee revenues result in a long-term sustainable APY.
where ˆαe is the recalibrated form ofαe appropriate for an update time step ∆t.
The above framework defines the token emission in the absence of any token supply flowing back into the network
treasury. In practice, all fees collected by the network are added to the network tre프로젝트로 돌아가기