Order-of-magnitude decision analysis describes each probability and utility only by its rough size, as a power of a small number $\epsilon$ ("likely," "rare," "very rare"). This lets an analyst build a sequential decision model, an influence diagram, which is a graph of chance events, decisions, and payoffs, without inventing precise numbers. The approach works well when every payoff is a gain. It breaks down when gains and losses of the same rough size meet: adding a "moderate gain" to a "moderate loss" yields a value the calculus can only call "moderate, sign unknown." That ambiguous value then spreads through the computation. Its effect is measurable. On mixed-sign problems, the typical policy that the qualitative solver returns shows high variance in quality, and no existing method fixes this.
We propose to stop treating that ambiguity as a defect of arithmetic and to treat it as what it usually is: incomparability between two criteria that the qualitative model never related to each other. Each value carries the size of its gains and the size of its losses as two separate coordinates. These pairs are compared by Pareto dominance, a partial order in which neither pair is better unless it is at least as good on both coordinates. A cancellation is then resolved by refining to the next power of $\epsilon$ only where an actual decision depends on it. The claim is that this removes most of the mixed-utility variance while keeping the small, bounded sets of candidate values that make order-of-magnitude elimination tractable.
Core Claim
The closest work is the order-of-magnitude influence diagram line of Marinescu and Wilson. Order-of-Magnitude Influence Diagrams represents values as pairs $(\sigma, n)$, a sign and an $\epsilon$-exponent, and solves the diagram with a bucket elimination algorithm called ELIM-OOM-ID. Bucket elimination removes variables one at a time by summing or maximizing them out. The paper shows that the best policy in the returned set is nearly exact. It also reports that the median policy shows higher variance, particularly on mixed-utility (Class M) problems, and attributes this to $(\sigma,n)\oplus(-\sigma,n)\to(\pm,n)$. An axiomatic framework for influence diagram computation with partially ordered preferences generalizes elimination to partially ordered utilities. It shows that distributivity fails in general but holds up to convex-closure equivalence, and that in the order-of-magnitude case any utility set compresses to at most two elements. That compression is exactly where gains and losses get merged. Both papers motivate this proposal and supply its machinery. Neither separates gains from losses, and neither offers any mechanism to bound or resolve the $\pm$ operator. That gap is ours.
Concretely, a value becomes $v=(g,\ell)$. Here $g$ is the $\epsilon$-exponent of the largest gain term and $\ell$ is the exponent of the largest loss term, with $\infty$ meaning "absent." Smaller exponents mean larger magnitudes. Addition acts componentwise, $(g_1,\ell_1)\oplus(g_2,\ell_2)=(\min(g_1,g_2),\min(\ell_1,\ell_2))$. Multiplying by a probability of order $\epsilon^k$ adds $k$ to both coordinates. Neither operation ever cancels anything. Dominance is
$$v_1 \succeq v_2 \iff g_1 \le g_2 \ \text{and}\ \ell_1 \ge \ell_2,$$
which says $v_1$ gains at least as much and loses at most as much. The old $(\pm,n)$ case becomes an explicit incomparable pair, and an incomparable pair is not allowed to absorb its neighbours. The undominated values of a finite set form an antichain in a two-dimensional grid of exponents. The set's size is therefore bounded by the number of distinct exponents $N$ that appear in the model, because an antichain in such a grid contains at most one element per gain-exponent. That bound is linear in $N$. It is not the 2-element bound, but it is still far from the unpredictable growth that general Pareto elimination suffers, which the axiomatic-framework paper identifies as its open problem.
The second ingredient is refinement on demand. At a decision node, two candidate options may stay incomparable only because their gain and loss exponents coincide. In that case we re-evaluate only the affected subtree, keeping one extra $\epsilon$ term in the style of the extended-real values of Qualitative MDPs and POMDPs: An Order-Of-Magnitude Approximation. Bonet and Pearl carry full rational functions of $\epsilon$ everywhere. We carry the extra term only where a decision is otherwise undetermined.
Research Program
The theoretical work has three parts. First, we will verify that the gain/loss structure satisfies the uncertainty-utility axioms of the axiomatic-framework paper, so that generalized bucket elimination remains correct up to convex-closure equivalence. Second, we will prove the antichain size bound. Third, we will characterize when refinement terminates: under exact rational inputs, a tie that persists at every order means the options truly have equal value.
The algorithmic work modifies ELIM-OOM-ID to propagate antichains of $(g,\ell)$ pairs. It adds a refinement trigger whenever a maximization over a decision variable leaves more than one undominated option. It also adds a budget on extra terms, so the cost of refinement can be measured.
The empirical work reuses the published benchmark generator exactly.
Conceptually, this is the poset direction acting as the core mechanism inside a qualitative decision procedure. Fargier and Perny's Qualitative Models for Decision Under Uncertainty without the Commensurability Assumption argues axiomatically that forcing two scales into one calibrated scale is unjustified. They stop at one-shot decisions with no algorithm. We apply the same refusal to commensurate gains with losses, inside a sequential solver.
Baselines
The naive reference is ELIM-OOM-ID exactly as published in Order-of-Magnitude Influence Diagrams, returning its qualitative optimal set with $\pm$ collapse. We report both the median policy and the best policy in that set, matching the published metrics. For the strongest reported number, An axiomatic framework for influence diagram computation with partially ordered preferences states that at $\epsilon=0.5$ the median relative error $\eta_{med}$ falls "below 10%" against the fully precise diagram, and that the best-policy error $\eta_{max}$ is "virtually zero." No reported result was found on Lacuna for the numerical variance of $\eta_{med}$ on Class M. We will therefore regenerate that baseline ourselves on identical instances.
The strongest existing method for this question is the generalized bucket elimination of the same axiomatic-framework paper, with gains and losses treated as two real-valued objectives under plain Pareto dominance. This approach avoids cancellation by design, so it is the true competitor. Its weakness is that its sets may grow without bound.
The ablations isolate our two ingredients. The first uses the split representation without refinement. The second uses ELIM-OOM-ID with refinement but without the split. In the second, refinement fires whenever a $\pm$ value appears, which is essentially a localized form of the Bonet–Pearl approach.
All methods run on the same generated instances with the same $\epsilon$ discretization of the same precise diagrams, and share one implementation of elimination order and memory limits. None has tunable hyperparameters beyond the refinement budget, which we sweep and report. We beat the baselines if Class M median error and its spread drop to Class P levels, while set sizes stay within a small constant of ELIM-OOM-ID and well below Pareto elimination at equal runtime.
Why not just…?
Why not just use a finer granularity such as $\epsilon=0.005$? A finer granularity makes distinct magnitudes less likely to share an exponent. It does not change the fact that equal-order gains and losses collapse to $\pm$, and the published variance problem persists across the tested granularities. Our experiments sweep all three published values of $\epsilon$. If finer $\epsilon$ alone closed the Class M gap, that would refute our motivation, and we would report it.
Why not just carry full rational functions of $\epsilon$, as Bonet and Pearl do? This is a real threat, because exact extended reals never cancel ambiguously. The cost is that values stop compressing to bounded canonical sets, and the analyst must supply the very coefficients that qualitative modelling exists to avoid. The refinement-only ablation is a localized version of this approach. If it matches the full method at similar cost, the split representation adds nothing, and the honest conclusion is that targeted refinement suffices.
Why not just treat gains and losses as two real objectives in Pareto elimination? This removes cancellation, but it uses real numbers, so frontier sizes are not tied to an exponent range. That is exactly the unpredictable growth that the axiomatic-framework paper reports for general partially ordered utilities. Running it as our strongest baseline settles whether the order-of-magnitude coordinates actually keep set sizes small.
Proposed Experiments
We will use the synthetic influence diagrams parameterized by $\langle n_c,n_d,k,p,r,a\rangle$, with chance variables from 15 to 60, both Class P and Class M, at $\epsilon\in\{0.5,0.05,0.005\}$, plus the Oil Wildcatter problem. These are the benchmarks of Order-of-Magnitude Influence Diagrams and An axiomatic framework for influence diagram computation with partially ordered preferences. We will record five metrics: $\eta_{med}$ and $\eta_{max}$ against the precise diagram, the interquartile spread of $\eta_{med}$ across instances, the maximum and mean canonical set size per bucket, and runtime.
The decisive check is the Class M minus Class P gap in median error and its spread. Under our account, the gap is caused by $\pm$ absorption, so the gap should shrink under the split representation while staying unchanged for ELIM-OOM-ID. A second check is diagnostic. We count the decisions at which ELIM-OOM-ID's optimal set contained a $\pm$-tainted value, and test whether these instances account for most of its median error. If they do not, cancellation is not the cause of the variance, and the proposal is wrong. Finally, we will check that measured set sizes never exceed the antichain bound, and how often refinement fires. Frequent firing would mean the qualitative abstraction is too coarse for mixed problems, which is itself a useful result.