Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | Let us look in more detail to the three objects detected in this study. 3C 459 shows a broad profile, at the resolution of our obser- vations, but we cannot be completely sure whether the absorption is against the nuclear region or the eastern lobe. Moreover, even assuming that the absorption is against the former, it appears that the arcsec nuclear region is actually formed by two steep spectrum components. Thus, VLBI observations will be required to clarify the real structure of the absorption. The absorption is not at the systemic velocity (although a more accurated redshift will be re- quired to confirm this) but appears to be blueshifted. Thus the torus, if present, is misaligned with respect to the radio axis (and this would explain why velocities different from the systemic are pro- jectd against the core). An alternative explanation is that an outflow of the neutral gas is present. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | velocity (assuming the new redshift reported in §3.4). Most of the H I absorption, including the very deep component, appears to be blueshifted. Thus, unless again the geometry of the disk respect to the radio axis is peculiar, or the systemic redshift substantially in error, this part of the absorption will not be easy to reconcile with a torus and could be instead gas associated with an expanding gas cocoon around the radio source. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | PKS 1549–79 seems to be a different kind of object. This galaxy has a core-jet radio structure and no counter-jet is observed at the resolution of the VLBI (≤ 10 mas, ∼ 35 pc, King et al.). The detected H I absorption has a lower column density and a nar- rower profile compared to the other two galaxies. With its core/jet VLBI morphology, it is difficult to reconcile the absorption as due to a torus: in that case the absorption would be expected against the counter-jet. The characteristics of the H I absorption in this object are more easy to reconcile with a cloud of neutral hydrogen in front of the nucleus. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | In summary, on one hand the fact that the H I absorption is detected in NLRGs and not in the BLRGs seems to suggest that a thick disk is indeed the cause of absorption. However, the above analysis shows that such a model does not explain all the features of the H I absorption in all NLRGs. The blueshifted absorption does indicate outflow and the possible correlation of the detection of H I absorption with the presence of a particularly rich ISM (see below) suggests that in some cases the absorption is related to an interac- tion between the radio plasma and the ISM. It is conceivable that this is connected to the fact that all the detected sources are com- pact or small. H I absorption related to gas outflow has been clearly seen already in other objects such as the Seyfert galaxy IC 5063 (Oosterloo et al. 2000), possibly the superluminal source 3C 216 (Pihlstr¨om et al. 1999) and 3C 326 (Conway et al. in prep). Thus, it is not unlikely that a similar situations can be present in some of our detected objects. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | As mentioned in §4, only two galaxies (3C 459 and PKS 1549–79) out of the 20 powerful and intermediate redshift (0.15 < z < 0.7) radio galaxies of the 2-Jy sample are dominated by a starburst at optical/UV wavelengths (Tadhunter 1999) and these two objects are the only two in the sample detected by IRAS at 60µm. It is therefore very interesting to see that in both these objects H I absorption has been detected. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | Although a contribution to the far-IR emission from quasar heating cannot be completely excluded, it is generally assumed that the far-IR excess represents reprocessed starlight. This points to a strong link between the star formation activity and the far-IR, and to an overall richer interstellar medium present in these galaxies. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | In order to test if the rich ISM could have some correlation with detection of H I absorption, we have searched the literature to investigate if this trend would hold for a larger number of objects. As far as we could find, three other starburst/far-IR bright galax- ies were observed in H I and all are detected in absorption. Apart from the already mentioned 4C 12.50 (Mirabel 1989), detections of H I absorption have been reported for 3C 433 (Mirabel 1989) and 3C 321 (Mirabel 1990). These all happen to be among the most powerful far-IR sources (for their range of radio power) and they all show strong signs of starburst activity in the optical. For ex- ample, 3C 433 is the object which shows the clearest evidence for optical/UV starburst activity and is also the most powerful object at far-IR wavelengths in a survey of a complete sample of 9 low red- shift (0.04 < z < 0.2) 3CR radio galaxies (Wills et al. in prep.). |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | Thus, although we do not have unbiased statistics, there seems to be an effect related to the presence of a rich ISM that makes the detection of H I in absorption more likely. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | As already pointed out by Mirabel (1982), there could be a link between radio-IR spirals (like the starburst Arp 220) and radio galaxies in the early stage of their evolution, like 4C 12.50. In the formation of radio galaxies and QSRs, the ultraluminous infrared galaxies (ULIRG) like Arp 220 are believed to play an important role (see Sanders and Mirabel 1996 for a review). They may repre- sent the early stage in a sequence of merging spiral galaxies evolv- ing through a starburst phase to eventually becoming QSOs and powerful radio galaxies. Thus, the starburst, far-IR galaxies that we are considering could represent the result of such a process. They should be either young radio galaxies, as suggested by the fact that many of them are compact and have a steep radio spec- trum, or galaxies with re-started activity (like 3C 321 which has a steep spectrum core or obvious mergers as 3C 433). A number of ULIRGs have been detected in H I absorption. Their typical H I col- umn density is > 1021−22 cm−2 for Tspin = 100 K. Their broad H I profiles indicate rotation plus large amounts of unusually turbulent neutral hydrogen (Mirabel 1982, Sanders & Mirabel 1996). |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | We have presented H I observations for 23 radio galaxies selected from the 2-Jy sample. H I absorption has been detected in 5 objects. We have investigated how the detection rate is distributed among galaxies with different optical and radio properties. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | For FR-I galaxies our data support the idea that most of the H I absorption in these objects comes from a thin nuclear disk, similar to those found in other galaxies of this type (NGC 4261 and Hy- dra A). Thus, this is consistent with the HST imaging observations (Chiaberge et al. 1999) which do not require the presence of pc scale, geometrically thick torus in low-luminosity radio galaxies. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | The non-detection of H I absorption in BLRGs is, to first order, consistent with the unified schemes. Broad line radio galaxies are in fact supposed to be galaxies seen pole-on and therefore obscuration from the torus should not occur. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | The fact that the H I absorption is detected in NLRGs and not in the BLRGs seems to suggest that a thick disk is indeed a cause of absorption. However, it does not seem to explain all the observed characteristics of the absorption. For example, the detection of H I absorption in far-IR bright, starburst galaxies seems to indicate that there is also an effect related to the presence of a rich ISM that makes detection of H I in absorption more likely. We cannot say if the H I is in a disk, but certainly in some objects the situation seems to be more complicated than that. In particular, PKS 1814–63 and 3C 459 seem to have blueshifted absorption, possibly related to some outflow. PKS 1549–79 is likely to be a very young radio source and the material around the nucleus has not yet been swept out by the expanding radio source and/or quasar-induced winds. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | It is clear that both high resolution VLBI data as well as ac- curate redshifts are required for a more complete picture of what is producing the H I absorption. |
Processed_HI_absorption_in_radio_galaxies:_effect_of_orienta.txt | We wish to acknowledge Rene Vermeulen and Ylva Pihlstr¨om for comunicating their results before publication. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The multi-armed bandit (MAB) framework has been intensively investigated in the last decade, handling the exploration vs exploitation dilemma through diverse selection rules, e.g. the upper confidence bound (UCB) criterion (Auer et al., 2002), KL-UCB (Maillard et al., 2011) or Thompson sampling (Chapelle and Li, 2011). The rise of MAB studies is explained as they tackle the rigorous analysis of algorithms in a both simplified and chal- lenging setting. On the one hand, the MAB framework defines a simplified reinforcement learning problem (Sutton and Barto, 1998; Szepesvari, 2010), where the state space involves a single state. On the other hand, MAB is concerned with lifelong learning, and the op- timization of the policy return while learning it. RL classically distinguishes the learning phase, and the production phase where the learned policy is applied. Most generally MAB makes no such distinction and aims at minimizing the cumulative regret suffered compared to the oracle strategy, during the whole learning/production life of the MAB system. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | due to mechanical fatigue and hazards. In the domain of energy management, simulators define noisy optimization problems as they must reflect the large variance of the energy load, demand and price along time; in the meanwhile, exploratory policies face huge losses if insufficient supply must be compensated for by buying additional energy. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | In risky environments, the goal becomes to learn a policy which achieves some trade- off between exploration, exploitation and safety: while the goal is to minimize the regret through a standard exploration vs exploitation trade-off, one also wants to minimize the risk incurred by the learned policy and/or to maximize the safety of the learning agent. Risk minimization is also enforced by considering short time horizons. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The importance of risk minimization, either within the MAB setting (Sani et al., 2012; Maillard, 2013) or generally in reinforcement learning (Moldovan and Abbeel, 2012b) has been recently acknowledged (section 2). A new algorithm, the multi-armed risk-aware ban- dit (MaRaB) algorithm, is presented in section 3. MaRaB aims at the arm with maximal conditional value at risk level α (CVaRα), where CVaRα is the expected policy return in the prescribed quantile. When α goes to 0, MaRaB tends toward the MIN multi-armed bandit algorithm, aimed at the arm with maximal minimal value. A theoretical analysis of the MIN algorithm shows that it achieves logarithmic regret under mild assumptions (section 4). Extensive empirical validation on artificial problems shows that MaRaB less explores the arms with low distribution tails compared to UCB and (Sani et al., 2012), at the expense of a moderate regret increase compared to UCB. A real-world problem related to battery management with a stochastic demand is also considered to investigate the robustness of the approach (section 5). The paper concludes with a discussion and some perspectives for further research. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | After introducing the multi-armed bandit (MAB) formal background and referring the reader to (Robbins, 1952; Auer et al., 2002) for a comprehensive presentation, this sec- tion briefly reviews the state of the art related to risk-aware MAB strategies. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | We will use the theoretical cumulative regret in the algorithm analysis (section 4). Theo- retical or empirical cumulative regrets will be used to experimentally assess the algorithms performance (section 5), granted that the difference |Rt − (cid:99)Rt| is in O(log(t)) (Coquelin and Munos, 2007). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Regret minimization is known to be an exploitation vs exploration trade-off problem: the best empirical arm should be selected often to maximize the actual gathered reward (exploitation); but some exploration is also required to actually identify the best arm. Two prominent MAB strategies are the (cid:15)-greedy strategy, which selects the best empirical arm with probability 1 − (cid:15) and uniformly selects another arm with probability (cid:15), and the cel- ebrated upper confidence bound (UCB) strategy proposed by Auer et al. (2002), which , with C > 0 a parameter selects at time t the arm maximizing criterion (cid:99)µi,t + C controlling the exploration vs exploitation tradeoff. Another strategy is the KL-UCB strat- egy (Maillard et al., 2011). While the (cid:15)-greedy strategy suffers a linear regret, UCB and KL-UCB suffer a logarithmic regret, which is known to be optimal (Lai and Robbins, 1985). KL-UCB further improves on UCB as it yields the optimal regret rate. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | An emerging trend in the field of reinforcement learning and MAB is concerned with the risk issue, when the exploring agent might face hazards going beyond mere under-optimal performances. In such cases, mottos such as Optimism in front of the Unknown! attached to the UCB strategy, are inappropriate. A first issue concerns the definition of risks. Sev- eral definitions have been proposed to account for risk awareness and risk aversion, taking inspiration from the literature in economics (Arrow, 1971). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The first criterion referred to as mean-variance (MV) (Markowitz, 1952) considers a weighted sum of the reward expectation µ of the policy and its estimated standard deviation. Formally, the goal is to find a policy minimizing σ2 − ρµ, where ρ > 0 increases like the user’s risk tolerance. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The conditional value at risk (CVaR) considers the quantiles of the reward distribution. Formally, let 0 < α < 1 be the target quantile level. The associated quantile value vα is defined if it exists as the maximal value such that X is less than vα with probability α (P r(X < vα) = α, with X the reward random variable). The remainder of the paper only considers continuous distributions, where vα is always defined; the conditional value at risk α noted CVaRα is then defined as the average reward conditionally to X < vα: CVaRα[X] = IE[X|X < vα]. Note that when the quantile level α goes to 0, CVaRα maximization coincides with the standard max-min strategy, aimed at the arm with maximal minimum reward. CVaR maximization thus defines a relaxation of the max-min strategy, with quantile level α as relaxation parameter. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | probabilities, e.g. the over-estimation of rare events, through weighting the event rewards with a (non-linear) function of their rank. The RDU criterion will not be considered further as it relies on a complex specification of the risk aversion, whereas the above MV and CVaR criteria involve a single scalar parameter, respectively ρ and α. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Risk issues have also been considered in the neighbor field of reinforcement learning in the last years. A first strategy relies on reversibility constraints, i.e. only visiting states s such that one can always get back from s to the initial state (Moldovan and Abbeel, 2012a). In a further work, Moldovan and Abbeel (2012b) proceed by considering an exponential utility function, where the policy return J is replaced by expression exp{J/θ}. Parameter θ reflects the user’s risk tolerance, akin the ρ parameter in the mean var setting, with the difference that ρ is weighted by the empirical standard deviation of the rewards. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Another approach due to Mannor and Tsitsiklis (2011) formalizes risk-aware reinforce- ment learning as a multi-objective RL problem, aimed at simultaneously maximizing the cumulative reward and minimizing the cumulative standard deviation. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | This section describes the Multi-Armed Risk-Aware Bandit (MaRaB) algorithm, with same notations as in section 2.1. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | MaRaB features a risk-averse or pessimistic behavior, due to the negative exploratory term in Eq. 2 : if two arms have same empirical CVaR, MaRaB will favor the arm which has been selected more often in the past. Note that such a behavior is actually observed in the economic realm, as trust − i.e. a positive bias toward known good partners − is at the core of economic exchanges. Such a bias indeed makes sense whenever exchanges with unknown partners involve risks. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | A lack of exploration usually leads to myopic and under-optimal choices, sticking to the best options first encountered. Such a myopic behavior is however prevented in MaRaB for the following reason: MaRaB examines each arm along two phases. In the first phase, referred to as initial phase (ni,t < 1 α ), the empirical quality of the i-th arm is set to its minimum reward (Eq. 1), and therefore it monotonically decreases along time. In the second phase, referred to as stabilization phase, the estimate of the conditional value at risk is computed with increasing accuracy, with an approximation error going to 0 like ni,t (Chen, 2008). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The duration of the initial phase increases as α decreases, as the maximization of the conditional value at risk α boils down to a standard max-min optimization problem. In the early iterations, the MaRaB behavior thus coincides with that of the MIN algorithm, selecting in each time step the arm with maximal minimum reward. The only difference comes from the negative exploration term (lower confidence bound, LCB1). MaRaB thus actually achieves some exploration in the beginnings: the arm quality (cid:92)CV aRi monotonically decreases as the i-th arm is more visited (ni,t increases) in its initial phase, forcing MaRaB to consider less visited arms. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | However, if an arm gets poor rewards the first times it is visited, there is little chance it is visited again, all the more so as better arms enter their second phases (and their empirical quality converges toward their true conditional value at risk): there is no positive exploration term guaranteeing that any arm will be visited infinitely many times as t goes to infinity. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | This section presents the analysis of the MIN algorithm, selecting in each time step the arm with maximal empirical minimal value, as MIN is the limit algorithm of MaRaB when the risk level α goes to 0 and the exploratory constant C is set to 0. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | with similar rate. Under slightly stronger assumptions, the MIN regret rate is significantly lower than for UCB. These two results rely on two lemmas. Firstly, under mild assumptions, the empirical minimum value for every arm converges exponentially fast toward its essential infimum. Secondly, with high probability over all arms their empirical minimum value are exponentially close to their essential minimum, where the probability increases exponentially fast with the number of iterations. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Under the assumption of a lower-bounded distribution probability in the neighborhood of its minimum, the convergence toward the minimum thus is faster than the convergence toward the mean. Specifically, the Hoeffding bound on the convergence toward the mean decreases exponentially like −t(cid:15)2, whereas after Eq. 4 the convergence toward the min de- creases exponentially like3 −tA(cid:15). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Where the first inequality follows from (1 − z)y ≥ 1 − y.z and the second inequality from (1 − z) ≤ exp(−z), which concludes the proof. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Under the above assumptions on the arm distributions, if the optimal arm in terms of min value also is the optimal arm in terms of mean value, then the MIN algorithm achieves a logarithmic regret. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | where i∗ stands for the index of the optimal arm. MIN and UCB thus both achieve a logarithmic regret uniformly over t, where the regret rate involves the mean-related margin in UCB (resp. the min-related margin in MIN, multiplied by the lower-bound constant A on the density in the neighborhood of the minimum). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Proposition 4.4 With same notations and assumptions as in Prop. 4.3, let us further assume that for every i = 1 . . . K, ∆µ,i = µ∗ − µi ≤ a∗ − ai = ∆a,i. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Discussion. The comparison of Eq. 13 and Eq. 12 suggests that MIN might outperform UCB in the case where margins ∆µ,i are small, where distributions νi are not too thin in the neighborhood of the essential infimum (that is, A is not too small), and the assumption ∆a,i ≥ ∆µ,i holds. Note that the latter assumption boils down to considering that better arms (in the sense of their mean) also have a narrower support for their lower tail, thus a lower risk. If this assumption does not hold however, then risk minimization and regret minimization are likely to be conflicting objectives. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | A last remark is that the assumptions done (lower bounded distribution density in the neighborhood of the essential minimum and mean-related margin greater than the minimum-related margin) yield a significant improvement compared to the continuous distribution-free case, where the optimal regret is known to be O( t) (Audibert and Bubeck, 2009, 2010). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | real-world problem in the target application domain of energy management is also considered (section 5.3). The goal of experiments is to answer three questions. The first one is the price to pay in terms of performance loss for a risk-aware behavior, and how the cumulative regret increases with the number of iterations, specifically focussing on short time horizons (unless explicitly specified, the empirical cumulative regret is considered). The second question regards the robustness of the algorithms, and their sensitivity w.r.t. parameters. A third question is whether MaRaB, MV-LCB and ExpExp do avoid exploring risky arms; this question is investigated by inspecting the low tail of the gathered rewards. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The number K of arms is set to 20. The time horizon is set to T = K × 100 and T = K × 200. For all problems, all results over (respectively the average result out of) 40 runs are displayed. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | An ad-hoc problem generator satisfying the assumptions done in Prop. 4.4 is used to compare MIN, UCB and MaRaB in the favorable case. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Figure 1: Theoretical cumulative regret of UCB, MIN and MaRaB under the assumptions of Prop. 4.4, averaged out of 40 runs. Parameter C ranges in {10i, i = −6 . . . 3}. Risk quantile level α ranges from .1% to 10%. Left: UCB regret increases logarithmically with the number of iterations for well-tuned C; MIN identifies the best arm after 50 iterations and its regret is constant thereafter. Right: zoom on the lower region of Left, with MIN and MaRaB regrets; MaRaB regret is close to that of MIN, irrespective of the C and α values in the considered ranges. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | theoretical cumulative regrets of UCB, MIN and MaRaB are displayed in Fig. 1 (averaged out of 40 independent runs with µ∗ = 0.5, a∗ = µ∗ − 10−3 and maximal radius 0.5). Param- eter C of UCB and MaRaB ranges in {10i, i = −6 . . . 3} and the risk level α ranges from .1% to 10%. By construction, this artificial problem favors MIN against UCB; firstly it satisfies the assumptions of Prop. 4.4; moreover since distributions νi are uniform, A ≥ 1. In this easy setting, MIN catches the best arm after 50 iterations and yields a constant regret thereafter (no exploration). MaRaB features the same behavior for a wide range of values of C and α; its very low sensitivity w.r.t. C slightly increases for high values of α (α > 20%). The disappointing UCB performance is blamed on the high variance of the worse arms, slowing down the accurate estimation of their mean. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | A second problem generator is considered, which only satisfies the assumption of a lower- bounded density in the neighborhood of the minimum (Eq. 3). Specifically, each problem involves 20 arms. The i-th arm distribution νi is set to a mixture of truncated Gaussians: i) its minimum ai is uniformly drawn in [0, .05]; ii) ni Gaussians are defined where ni is uniformly drawn in 1 . . . 4; for j = 1 . . . ni the j-th Gaussian N (µi,j, σi,j), is defined by uniformly sampling µi,j in [0, 1] and σi,j in [.12, .5]; furthermore, the j-th Gaussian is associated a probability pi,j such that (cid:80) j pi,j = 1. Upon selecting the i-th arm, the reward is drawn by: i) selecting the j-th Gaussian with probability pi,j; ii) drawing a reward r from N (µi,j, σi,j); iii) going to i) if r is not in the [ai, 1] interval (rejection-based truncation). |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The empirical cumulative regrets of UCB, MaRaB, MV-LCB and ExpExp are displayed in Fig. 2, reporting the empirical cdf5 of the regrets over 1,000 problem instances for short time (Fig. 2.(a)) and medium time (Fig. 2.(b)) horizons. All algorithm parameters are set to their best value after preliminary experiments. UCB yields the best cumulative regret overall whenever C is well tuned. MaRaB suffers an extra regret compared to UCB; this extra regret is bounded in the considered experimental setting, and it seemingly does not increase as the time horizon increases. As could have been expected this extra regret decreases as α increases and the selection rule involves a better estimation of the empirical means. Interestingly, MaRaB shows a very low sensitivity w.r.t. C. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | ri is an increasing affine function of i, with r1 = µ∗ − a∗ and ri = r1 + i−1 The mean-related margin ∆µ,i is thus controlled from ∆max; the min-related margin ∆a,i is controlled from ∆max and rmax, in such a way that ∆a,i > ∆µ,i. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | improves on ExpExp with probability 70%, albeit with maximal cumulative regrets (over the problem instances) higher than for ExpExp. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Overall, MaRaB with risk level α = 20% and untuned C value yields results slightly less than UCB with tuned C, for both short and medium time horizons. The risk-aware MaRaB suffers a low regret increase compared to risk-neutral UCB, with a very low sensitivity w.r.t. C. Interestingly, a twice longer time horizon does not modify the performance order of the algorithms. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The effective risk avoidance of UCB, MV-LCB, ExpExp and MaRaB are investigated by inspecting the empirical cdf6 of the instant rewards on two representative artificial problems, with respectively low (Fig. 3, left) and high (Fig. 3, right) variance of the best arm. The low tail of the cdf (worst average rewards gathered by the algorithm) indicates whether the algorithm actually tried poor arms. Fig. 3 confirms previous results: The noted sensitivity of UCB w.r.t. parameter C unsurprisingly increases with the variance of the best arm (Fig. 3, top row). The bad performance of MV-LCB is confirmed; its sensitivity w.r.t. ρ is low on the low variance problem as expected (Fig. 3, second row, left); its sensitivity w.r.t. ρ is much higher on the high variance problem (Fig. 3, second row, right), with a best performance for medium values of ρ. ExpExp features an excellent risk avoidance as the risky trials only take place during the exploratory phase (Fig. 3, third row). The general robustness of MaRaB w.r.t. C is confirmed; moreover, its robustness w.r.t. the risk level α on high variance problems is empirically shown (Fig. 3, bottom row). It is seen that for low to medium risk (α < 20%), the quantile values vα (section 2.2) are consistently higher for MaRaB than for ExpExp, which is explained again from the systematic exploratory phase in ExpExp. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Figure 3: Comparative risk avoidance of UCB, MV-LCB, ExpExp and MaRaB on two representative artificial problems with low (left) and high (right) variance of the optimal arm. For each algorithm instant rewards averaged out of 40 runs are sorted. The time horizon is set to T = 2, 000. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The real-world problem motivating the presented approach is a battery management prob- lem, where the environment is described by the energy demand and the energy cost in each time step. The decision to be taken in each time step is a real-value x, determining how much energy is either used from the battery (if x > 0) or stored in the battery (if x < 0). In each time step, one must meet the demand by buying min(0, demand − x) energy; the instant reward is the cost of the bought energy if the demand exceeds the available en- ergy. Additionally, the battery loses some energy in each time step. A simplified setting is considered, where i) the energy cost is constant, the random process only dictates the energy demand in each time step; ii) 20 arms, corresponding to pre-defined strategies are considered. The strategy reward is drawn by uniform sampling with replacement from the 117 available realizations of the strategy. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Same general trends as for the artificial problems are observed on this real-world problem i) The cumulative regret is minimal for UCB with optimally tuned C; ii) MV- (Fig. 4): LCB is dominated by all other algorithms w.r.t. both risk avoidance and cumulative regret; iii) the ExpExp regret increases linearly during the exploration phase and then reaches a plateau; iv) MaRaB shows its good risk-avoidance ability regardless of the C value, and MIN yields same results. Overall, MaRaB suffers a slight regret increase compared to UCB at its best, with a slightly better reward cdf in the region of low rewards. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | The first contribution of the paper, as a step toward an effective trade-off between exploita- tion, exploration and safety, is to show the theoretical soundness of the MIN algorithm. i) same arms are optimal in the perspective This result relies on two main assumptions: of regret and risk minimization; ii) the arm reward distributions are lower bounded in the neighborhood of their minimum on the other hand. Not only does MIN achieve logarith- mic regret; it also yields a better rate than UCB under the additional assumption that min-related margins are higher than mean-related margins. A second contribution is the MaRaB strategy, yielding a reduced risk at the expense of a moderate regret increase compared to UCB for short and medium time horizons, on artifial problems (which only satisfies the lower-bounded distribution assumption) and on a real-world one. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Further work is concerned with the analysis of MaRaB behavior, specifically its mean- related regret (under the assumption that the arm with best mean also is the arm with best CVaR) and its CVaR-related regret; another priority is to compare MaRaB behavior with that of Maillard (2013). A second perspective regards the case where the arms belong to a metric space; the goal becomes to exploit this metrix to enforce exploration safety. Last but not least, MaRaB will be extended to tree-structured search spaces to achieve e.g. safe sequential decision making. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | We are grateful to J.-J. Christophe, J. Decock and the members of the Ilab Metis and Artelys, for fruitful collaboration. We thank the anonymous referees for their insightful comments. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Figure 4: Comparative performance of UCB, MV-LCB, ExpExpand MaRaB on a real- world energy management problem. Left: sorted instant rewards (truncated to the 37.5% worst cases for readability). Right: empirical cumulative regret with time horizon T = 100K, averaged out of 40 runs. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | K.J. Arrow. Essays in the Theory of Risk-Bearing, chapter The Theory of Risk Aversion, pages 90–120. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Odalric-Ambrym Maillard, R´emi Munos, and Gilles Stoltz. A finite-time analysis of multi-armed bandits problems with Kullback-Leibler divergences. Journal of Machine Learning Research - Proceedings Track, 19:497–514, 2011. |
Processed_Exploration_vs_Exploitation_vs_Safety:_Risk-averse.txt | Teodor Mihai Moldovan and Pieter Abbeel. Safe exploration in Markov decision processes. In ICML, 2012a. |
Processed_Disentangling_the_Cosmic_Web_Towards_FRB_190608.txt | The Fast Radio Burst (FRB) 190608 was detected by the Australian Square-Kilometer Array Pathfinder (ASKAP) and localized to a spiral galaxy at zhost = 0.11778 in the Sloan Digital Sky Survey (SDSS) footprint. The burst has a large dispersion measure (DMFRB = 339.8 pc cm−3) compared to the expected cosmic average at its redshift. It also has a large rotation measure (RMFRB = 353 rad m−2) and scattering timescale (τ = 3.3 ms at 1.28 GHz). Chittidi et al. (2020) perform a detailed analysis of the ultraviolet and optical emission of the host galaxy and estimate the host DM contribution to be 110 ± 37 pc cm−3. This work complements theirs and reports the analysis of the optical data of galaxies in the foreground of FRB 190608 in order to explore their contributions to the FRB signal. Together, the two studies delineate an observationally driven, end-to-end study of matter distribution along an FRB sightline; the first study of its kind. Combining our Keck Cosmic Web Imager (KCWI) observations and public SDSS data, we estimate the expected cosmic dispersion measure DMcosmic along the sightline to FRB 190608. We first estimate the contribution of hot, ionized gas in interven- ing virialized halos (DMhalos ≈ 7 − 28 pc cm−3). Then, using the Monte Carlo Physarum Machine (MCPM) methodology, we produce a 3D map of ionized gas in cosmic web filaments and compute the DM contribution from matter outside halos (DMIGM ≈ 91 − 126 pc cm−3). This implies a greater frac- tion of ionized gas along this sightline is extant outside virialized halos. We also investigate whether the intervening halos can account for the large FRB rotation measure and pulse width and conclude that it is implausible. Both the pulse broadening and the large Faraday rotation likely arise from the progenitor environment or the host galaxy. |
Processed_Disentangling_the_Cosmic_Web_Towards_FRB_190608.txt | tion process of galaxies is further predicted, at least for halo masses Mhalo (cid:38) 1012 M(cid:12), to generate a halo of baryons, most likely dominated by gas shock-heated to the virial temperature of the potential well (White & Rees 1978; White & Frenk 1991; Kauffmann et al. 1993; Somerville & Primack 1999; Cole et al. 2000). At T (cid:38) 106 K and ne ∼ 10−4 cm−3, however, this halo gas is very difficult to detect in emission (Kuntz & Snow- den 2000; Yoshino et al. 2009; Henley & Shelton 2013) and similarly challenging to observe in absorption (e.g. Burchett et al. 2019). And while experiments leveraging the Sunyaev-Zeldovich effect are promising (Planck Col- laboration et al. 2016a), these are currently limited to massive halos and are subject to significant systematic effects (Lim et al. 2020). |
Processed_Disentangling_the_Cosmic_Web_Towards_FRB_190608.txt | Therefore, there has been a wide range of predictions for the mass fraction of baryons in massive halos that range from ≈ 10% to nearly the full complement rela- tive to the cosmic mean Ωb/Ωm (Pillepich et al. 2018). Here, Ωb and Ωm are the average cosmic densities of baryons and matter respectively. Underlying this order- of-magnitude spread in predictions are uncertain physi- cal processes that eject gas from galaxies and can greatly shape them and their environments (e.g. Suresh et al. 2015). |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Ground vehicles equipped with monocular vision systems are a valuable source of high reso- lution image data for precision agriculture applications in orchards. This paper presents an image processing framework for fruit detection and counting using orchard image data. A general purpose image segmentation approach is used, including two feature learning algo- rithms; multi-scale Multi-Layered Perceptrons (MLP) and Convolutional Neural Networks (CNN). These networks were extended by including contextual information about how the image data was captured (metadata), which correlates with some of the appearance vari- ations and/or class distributions observed in the data. The pixel-wise fruit segmentation output is processed using the Watershed Segmentation (WS) and Circular Hough Transform (CHT) algorithms to detect and count individual fruits. Experiments were conducted in a commercial apple orchard near Melbourne, Australia. The results show an improvement in fruit segmentation performance with the inclusion of metadata on the previously bench- marked MLP network. We extend this work with CNNs, bringing agrovision closer to the state-of-the-art in computer vision, where although metadata had negligible influence, the best pixel-wise F1-score of 0.791 was achieved. The WS algorithm produced the best apple detection and counting results, with a detection F1-score of 0.858. As a final step, image fruit counts were accumulated over multiple rows at the orchard and compared against the post-harvest fruit counts that were obtained from a grading and counting machine. The count estimates using CNN and WS resulted in the best performance for this dataset, with a squared correlation coefficient of r2 = 0.826. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Yield estimation and mapping in orchards is important for growers as it facilitates efficient utilisation of resources and improves returns per unit area and time. With accurate knowledge of yield distribution and quantity, a grower can efficiently manage processes such as chemigation, fertigation and thinning. Yield estimation also allows the grower to plan ahead of time their harvest logistics, crop storage and sales. The standard approach to get yield information is currently manual sampling, which is labour intensive, expensive and often destructive. Constrained by these costs, sampling is often done over a few individual crops and the measures are extrapolated over the entire farm. Inherent human sampling bias and sparsity in the measurements can result in inaccurate yield estimation. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | with standard colour cameras to capture a detailed representation over large farms. Robust and accurate image processing techniques are required to extract high level information from this data such as crop location, health, maturity, crop load (yield), spatial distribution etc. Image segmentation is the process of evaluating a semantic description of an image at a pixel or super-pixel level. For image data captured at an orchard, this means automatically labelling each pixel, or groups of pixels as representing fruits, flowers, trunks, branches and foliage. The parsed information can then be used in higher level tasks such as individual fruit detection, crop health analysis and tree detection/branch modelling. This provides the grower with a rich farm inventory and enables further robotic operations such as autonomous pruning, harvesting, variable rate spraying etc. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | A standard approach for image segmentation is to transform image regions into discriminative feature repre- sentations and parse them through a trained classifier, assigning each region a specific label. For identifying orchard fruits this means extraction of a feature space that captures properties unique to the fruit such as its colour, texture, reflection, position, shape etc. Farm image data is generally prone to a wide range of intra-class (within class) variations due to illumination conditions, occlusions, clustering, camera view-point, tree types, seasonal maturity (translating to a different fruit size, shape and colour) etc. Therefore, the feature space and the classifier need to be invariant to such characteristics. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Typically, prior works utilise hand engineered features to encode visual attributes that discriminate fruits from non-fruit regions (Payne et al., 2014; Wang et al., 2013; Nuske et al., 2014). Although these approaches are well suited for the dataset they are designed for, the feature encoding is generally unique to a specific fruit and the conditions under which the data were captured. As a result, the methods are often not transferable to other crops/datasets. In contrast, supervised feature learning approaches can be used to automatically learn transformations that capture the data distribution, enabling their use with different datasets (Hung et al., 2013). Such approaches have high model complexity and utilise extensive training exemplars to model the variations in the data (Krizhevsky et al., 2012). |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Image classification algorithms have also been shown to benefit from prior knowledge about scene structure. For example, Tighe and Lazebnik (2013) and Brust et al. (2015) specify a spatial prior over the labelled classes to aid image segmentation of public image datasets such as LabelMeFacade and SIFT Flow. A greater wealth of prior knowledge is often available in field robotics applications, as we often have access to contextual information about how the data were captured. For a typical image dataset, such contextual information, which we refer to as metadata, can include camera trajectories, vehicle location, type of tree/fruit being scanned, distance from camera to trees, sun position, illumination incident angle, weather conditions, etc. While a direct physical model of how particular meta-parameters affect the data is not available, aspects of the relationship between metadata and object classes can be learnt. Where there is correlation between metadata and appearance variations and/or class distributions, including metadata can improve classification performance. Available at no extra cost to typical data capturing process, our previous works (Bargoti and Underwood, 2015, 2016) have illustrated the use of metadata in allowing simpler classifiers to capture that space, and provide a performance boost, leading to similar performance with reduced training exemplars. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Figure 1: Left: The research ground vehicle Shrimp traversing between rows at an apple orchard, capturing tree image data. Location of apples manually illustrated in the field of view of the camera. Right: Satellite view of the 0.5 ha orchard test block. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Image fruit segmentation analysis using the previously benchmarked multi-scale Multi-Layered Per- ceptron (Hung et al., 2013; Bargoti and Underwood, 2016) and the state-of-the-art Convolutional Neural Networks. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | A study of the utility of different metadata sources and their inclusion within the different classifi- cation architectures and training configurations. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | The remainder of this paper is organized as follows. Section 2 presents related work on image classification in outdoor scenes, both in the agricultural and the general computer vision context. The image processing components are presented over Sections 3 to 6, following the computation pipeline illustrated in Figure 2. Section 3 presents the different classification frameworks with the inclusion of additional metadata. In Section 4 we outline the experimental setup and the image dataset, followed by Section 5 presenting the image segmentation results. Section 6 focuses on fruit detection and yield estimation performed over the image segmentation output. We present the practical lessons learnt in Section 7, including new insights into image processing for orchard data. We conclude in Section 8, discussing the future directions of this work. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Computer vision in agriculture, agrovision (Kapach et al., 2012), has been explored in multiple literature studies for the purposes of fruit detection and yield estimation (Jimenez et al., 2000; Payne and Walsh, 2014; Kapach et al., 2012). Agrovision literature is typically data specific, designed for the task at hand, and can often be very heuristic when compared against the most recent work in the general computer vision community (Kapach et al., 2012). In this section we discuss the key approaches used in agrovision while comparing them against the state-of-the-art techniques used in computer vision. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Image processing at orchards spans a large variety of fruits such as grapes (Nuske et al., 2014; Font et al., 2015), mangoes (Chhabra et al., 2012; Payne et al., 2014), apples (Linker et al., 2012; Wang et al., 2013; Stajnko et al., 2009; Silwal et al., 2014; Ji et al., 2012; Kim et al., 2015; Hung et al., 2015), citrus (Li et al., 2011; Annamalai et al., 2004; Sengupta and Lee, 2014; Regunathan and Lee, 2005; Qiang et al., 2014), kiwifruit (Wijethunga et al., 2009) and peaches (Kurtulmus et al., 2014). Fruit classification is generally performed by transforming image regions into discriminative feature spaces and using a trained classifier to associate them to either fruit regions or background objects such as foliage, branches, ground etc. If conducted densely, image regions are contextual windows neighbouring every pixel in the image and the output is a densely segmented image. Post-processing techniques can then be applied to differentiate individual objects of interest. A detection specific approach on the other hand, reduces the region search space by initially performing key-point detection. Here interesting image regions (possible fruit candidates) are first extracted using manually tuned constraints designed for the particular dataset. This is followed by feature extraction and classification as before. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Figure 2: The image segmentation, fruit detection and yield mapping pipeline. An image segmentation model is trained using labelled images extracted sparsely from the farm. This is used to segment the dense image data captured by a UGV. Detection algorithms are applied to the segmentation output to identify individual apples and accumulate the counts over the farm. The result is a dense farm yield map, which can be calibrated with ground truth counts to provide a yield estimate. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | use simple colour classifiers for key-point extraction for grape bunches and peppers respectively. For fruit detection, image patches are extracted around each key-point and a combination of colour and texture filters are computed. The patches can then be classified as fruits or not-fruits using a trained classifier such as a Support Vector Machine (SVM) or a randomised KD-forest. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Image segmentation on the other hand returns a rich likelihood map of the fruits, onto which a threshold can be applied to obtain a binary fruit mask detailing regions of the image containing fruit (Payne et al., 2014; Linker et al., 2012; Yamamoto et al., 2014; Sa et al., 2015). Payne et al. (2014) designs a set of heuristic measures based on local colours and textures to classify individual pixels as mangoes or non- mangoes. Blob extraction was done on the resultant binary mask to identify individual mangoes. Linker et al. (2012) incorporates further post-processing for apple detection where individual blobs are expanded, segmented and combined to manage occluded fruits and fruit clusters. Stajnko et al. (2009) instead uses shape analysis and template matching to extract circular apples from the segmented image. Yamamoto et al. (2014) implements a second classification component on tomato blobs extracted via image segmentation to remove any background detections. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Typically, orchard image data is subject to highly variable illumination conditions, shadowing effects, fruits/crops of different shapes and sizes, captured over different seasons, etc (Payne and Walsh, 2014), which makes classification a challenging task. To simplify and minimise the variations in the data, one can enforce constraints on the environment or the data gathering operation. For example, pepper detection in Song et al. (2014) is conducted in a greenhouse with controlled illumination conditions. Equivalently, in Nuske et al. (2014); Payne et al. (2014); Font et al. (2015) the data is captured at night using strobes, which significantly restricts the illumination variance. However, in orchards it is generally more practical to conduct large scale experiments under natural day-light conditions, and for commercial applications simple hardware such as cameras can be easily incorporated onto tractors, which operate more frequently during the day1. Therefore, image classification under natural illumination conditions is an open and important problem. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | As stated in the previous section, hand engineered feature encoding often restricts methods to particular fruits/datasets as they are designed to capture data specific inter-class variations (i.e. differences between trees, leaves and fruits), while being invariant to the intra-class variations. Although the methods stated above have produced promising performance over the respective fruits/datasets, they are distinct and ad hoc, seldom replicated, and often disconnected from progress in the general computer vision literature (Kapach et al., 2012). For widespread application, it would be more efficient to have unified image processing approach compatible with different fruits or capturing configurations. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | A general purpose adaptive feature learning algorithm is therefore desirable, as proposed in (Hung et al., 2013). Here, a pixel-wise image segmentation framework was presented, that utilises a multi-scale feature learning architecture to learn relevant feature transformations for pixel-level classification. The proposed approach was shown to outperform classification using hand engineered features. Additionally, the same architecture has been used for different datasets such as almonds (Hung et al., 2013), apples (Hung et al., 2015; Bargoti and Underwood, 2015, 2016) and tree trunks (Bargoti et al., 2015) without any changes to the image segmentation pipeline. However, even though such a feature learning approach avoids the need for manually designed features, there are opportunities for significant improvements, particularly to address the reliability in adverse illumination conditions, such as underexposed and overexposed images, where performance was observed to deteriorate. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Networks (CNNs) for image segmentation, where individual patches representing contextual regions around pixels are densely classified in an image. More recently, CNNs have been shown to yield improved seg- mentation performance when a spatial prior on the classes is available. Brust et al. (2015) perform road image segmentation while incorporating the pixel position, to help the classifier learn that road pixels are pre-dominantly found near the bottom half of images. |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | In this paper we evaluate the performance of a previously benchmarked multi-scale Multi-Layered Perceptron (MLP) architecture for image segmentation in agrovision (Hung et al., 2013) including an extension with metadata, which we have shown improves performance significantly in Bargoti and Underwood (2015, 2016). We further extend this study with comparison against state-of-the-art CNNs, with and without the addition of metadata, showing improved pixel classification performance. We evaluate the utility of improved image segmentation towards fruit detection and yield estimation. With this we also shorten the gap between image processing techniques used in agrovision to the current work in computer vision literature, which is a limitation addressed in the literature survey conducted in Kapach et al. (2012). |
Processed_Image_Segmentation_for_Fruit_Detection_and_Yield_E.txt | Image segmentation is the task of transforming individual pixels in an image into class labels. In this paper we present multiple image segmentation architectures for the binary classification of orchard image data into fruit/non-fruit classes. These include a multi-scale MLP and a state-of-the-art deep CNN architecture. We then extend these neural network architectures with the inclusion of orchard metadata. All network training is done at the pixel level, however inference is performed over the whole image, resulting in a dense probabilistic output of the fruit and non-fruit classes. This can be used to obtain a binary fruit mask and subsequently perform fruit detection or yield estimation as detailed in Section 6. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | Given the efficiency and equity concerns of a cordon toll, this paper proposes a few alternative distance-dependent area-based pricing models for a large-scale dynamic traffic network. We use the Network Fundamental Diagram (NFD) to monitor the network traffic state over time and consider different trip lengths in the toll calculation. The first model is a distance toll that is linearly related to the distance traveled within the cordon. The second model is an improved joint distance and time toll (JDTT) whereby users are charged jointly in proportion to the dis- tance traveled and time spent within the cordon. The third model is a further improved joint distance and delay toll (JDDT) which replaces the time toll in the JDTT with a delay toll com- ponent. To solve the optimal toll level problem, we develop a simulation-based optimization (SBO) framework. Specifically, we propose a simultaneous approach and a sequential ap- proach, respectively, based on the proportional-integral (PI) feedback controller to iteratively adjust the JDTT and JDDT, and use a calibrated large-scale simulation-based dynamic traffic assignment (DTA) model of Melbourne, Australia to evaluate the network performance under different pricing scenarios. While the framework is developed for static pricing, we show that it can be easily extended to solve time-dependent pricing by using multiple PI controllers. Re- sults show that although the distance toll keeps the network from entering the congested regime of the NFD, it naturally drives users into the shortest paths within the cordon resulting in an uneven distribution of congestion. This is reflected by a large clockwise hysteresis loop in the NFD. In contrast, both the JDTT and JDDT reduce the size of the hysteresis loop while achiev- ing the same control objective. We further conduct multiple simulation runs with different ran- dom seed numbers to demonstrate the effectiveness of different pricing models against simu- lation stochasticity. However, we postulate that the feedback control is not applicable with guaranteed convergence if the periphery of the cordon area becomes highly congested or grid- locked. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | launched the world’s first pricing scheme called the area licensing scheme. Following Singa- pore’s success, several other pricing schemes have been implemented worldwide, such as in London, Stockholm, and Milan. See Gu et al. (2018) for a comprehensive overview. More recently, instead of looking into cordon and zonal schemes, a number of advanced pricing schemes have been proposed and implemented thanks to the development of various pricing technologies. For example, (i) the Move NY Plan in New York City which aims to charge taxis based on both distance and time1; (ii) Singapore’s ERP 2.0 (from 2020) which uses satellites to charge vehicles based on distance2 traveled; (iii) the opt-in user-pays system OReGo in Or- egon, USA which is a state-wide distance-based charge3; (iv) a per-km charging system con- sidered by the Metro Vancouver Independent Mobility Pricing Commission (Lee, 2018); and (v) a joint distance- and cordon-based scheme trialed in Melbourne (Transurban, 2016). Inter- ested readers may refer to De Palma and Lindsey (2011) for an overview of different technol- ogies applied to pricing. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | In theory, since the seminal studies by Pigou (1920) and Knight (1924), a variety of mod- els have been proposed for designing an optimal pricing system. These models, however, can- not be easily applied to a large-scale dynamic traffic network because of their demanding re- quirement of highly detailed information of each origin-destination (OD) pair and each indi- vidual link in the network. See Yang and Huang (2005) for a comprehensive overview. Recent advances in network traffic flow theory through the Network Fundamental Diagram (NFD) or Macroscopic Fundamental Diagram (MFD) (Geroliminis and Daganzo, 2008; Mahmassani et al., 2013) have established a new branch of pricing theory that largely facilitates the study, design, and implementation of large-scale pricing models. Nevertheless, studies so far are mainly limited to the cordon-based regime which, although being simple, suffers from ineffi- ciency and inequity. Therefore, this paper aims to propose a few alternative distance-dependent area-based pricing models for large-scale dynamic traffic networks. To find the optimal toll levels that keep the network from entering the congested regime of the NFD, we develop a simulation-based optimization (SBO) framework combining the NFD and a simulation-based dynamic traffic assignment (DTA) model. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | any exact solution algorithm (Chen et al., 2016; Chen et al., 2014; Ekström et al., 2016; He et al., 2017). |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | To develop an effective and efficient pricing model for large-scale dynamic traffic net- works, an understanding of traffic dynamics at the network level is critical (Zheng et al., 2016). Macroscopic traffic flow relations for an urban traffic network were initially proposed by Godfrey (1969) followed by Daganzo (2007); Mahmassani et al. (1987); Olszewski et al. (1995). The formal demonstration of the existence of the NFD was shown only recently using field data from Yokohama, Japan (Geroliminis and Daganzo, 2008). Since then, the NFD has been widely studied for network-wide control and management (Geroliminis et al., 2013; Keyvan-Ekbatani et al., 2012; Ramezani et al., 2015) due to its macroscopic nature and favor- able properties (see Geroliminis and Daganzo (2008) for further description). In particular, an innovative NFD-based approach to modeling congestion pricing has emerged. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | Since macroscopic modeling does not involve detailed route choice, an urban traffic net- work governed by an NFD can be modeled as a simple queuing system (e.g. a bottleneck) represented by the cumulative arrival and exit curves. Congestion is interpreted as vehicles queuing behind the bottleneck while demand results from their endogenous scheduling prefer- ences considering the discrepancy between the desired and actual arrival times (Arnott et al., 1990; Arnott et al., 1993; Vickrey, 1963). Geroliminis and Levinson (2009) first integrated the bottleneck model with the NFD to consider the capacity drop. Daganzo and Lehe (2015) further introduced trip length heterogeneity and proposed an optimal usage-based toll. A fundamental study on NFD-based pricing in large-scale real-world networks was only recently conducted where the NFD was used to describe the level of congestion at the network level (Zheng et al., 2012). The authors applied an integral (I-type) controller for toll adjustment within an agent- based simulation environment. The I-type controller was later extended to a proportional-inte- gral (PI) controller whereby users’ adaptation to pricing was considered (Zheng et al., 2016). A similar simulation framework based on the NFD was also proposed by Simoni et al. (2015) but without using feedback control theory. The authors developed the toll adjustment rule by integrating the NFD (and also the 3D-NFD) with the MCP. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | NFD-based pricing in large-scale networks has so far mainly focused on the cordon-based regime. Given that a cordon toll undercharges long journeys and overcharges short trips, a distance toll was considered as an alternative (May and Milne, 2000; Meng et al., 2012). Alt- hough theory on distance-based pricing is well established, it is largely constrained within the traditional second-best pricing framework. Investigating distance-dependent pricing using the NFD therefore becomes theoretically interesting and also practically promising. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | The overall toll design problem consists of a toll level problem and a toll area problem (Ekström et al., 2012). In this paper, we assume a predefined toll area and only aim to solve the optimal toll level problem. Since we are solving an area-based pricing problem, the NFD becomes a natural entry point to the solution due to its macroscopic nature. Similar to the ex- isting studies on NFD-based network-wide control and management (Aboudolas and Geroliminis, 2013; Keyvan-Ekbatani et al., 2012; Simoni et al., 2015; Zheng et al., 2016; Zheng et al., 2012), the objective is to price and keep the network from entering the congested regime of the NFD. This is because when the network becomes congested or gridlocked (i.e. the net- work density increases beyond the critical threshold), the outflow/trip completion rate signifi- cantly drops causing network unproductivity. We assume that the NFD does not change sig- nificantly when we implement different pricing strategies on the network. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | networks. While using the NFD to describe the level of congestion at the network level, we consider different trip lengths in the toll calculation. The first model is a distance toll that is linearly related to the distance traveled within the cordon. A limitation, however, is that it nat- urally drives users into the shortest paths within the cordon resulting in an uneven distribution of congestion. The increased spatial traffic heterogeneity results in clusters of congested links within the cordon (i.e. pockets of congestion) and therefore (i) affects the shape of the NFD and deteriorates the network performance (Mazloumian et al., 2010; Saberi and Mahmassani, 2012, 2013); and (ii) compromises the toll optimality (Simoni et al., 2015). To address the limitation of the distance toll, we propose a joint distance and time toll (JDTT) as the second model based on Liu et al. (2014). The JDTT is linearly related to both the distance traveled and time spent within the cordon. Since the time toll component tends to overcharge users (because a longer link typically requires more travel time despite being uncongested), we further propose a joint distance and delay toll (JDDT) as the third model whereby users pay partially in pro- portion to their travel delays. To solve the optimal toll level problem, we develop an SBO framework which resembles, to some extent, the trial-and-error method in Yang et al. (2004). Specifically, we apply a PI feedback controller for iterative toll adjustment and use a calibrated large-scale simulation-based DTA model of Melbourne, Australia to evaluate different pricing scenarios. Practical applications of this type of iterative pricing strategy can be found in Sin- gapore (Liu et al., 2013) and San Francisco (Pierce and Shoup, 2013). While the framework is developed for static pricing, we show that it can be easily extended to solve time-dependent pricing by using multiple PI controllers. That is, we implement a PI controller in each tolling interval instead of using only one for the entire tolling period. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | i. We show that the distance toll naturally drives users into the shortest paths within the cordon resulting in an uneven distribution of congestion and hence creating a large clockwise hysteresis loop in the NFD. To reduce the increased spatial traffic heterogeneity resulting from the distance toll, we integrate a time toll component and a delay toll component, respectively, in the toll calculation resulting in two improved area-based pricing models: JDTT and JDDT. To solve the optimal toll level problem, we develop an SBO framework and propose a simultaneous approach and a sequential approach, respectively, based on the PI con- troller for iteratively adjusting the JDTT and JDDT. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | iv. We postulate a prerequisite for applying the PI controller: the periphery of the cordon area should have sufficient capacity to accommodate the re-routed traffic. If the prereq- uisite is not satisfied, the pricing control may fail resulting in a paradox where the toll levels keep rising but users still enter the cordon. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | The remainder of the paper is organized as follows. Section 2 proposes the methodological framework for solving the optimal toll level problem. Section 3 presents a case study where the network performance under different pricing scenarios is discussed and compared. Section 4 postulates a prerequisite for applying the feedback control. Section 5 concludes the paper and makes recommendations for future research. |
Processed_Optimal_distance-_and_time-dependent_area-based_pr.txt | Under the assumption that the NFD does not change significantly due to the imposition of different tolls, we propose our methodological framework consisting of the following four parts. Section 2.1 describes the construction of multiple macroscopic traffic flow variables us- ing simulation outputs. Section 2.2 introduces the route choice model that is used in simulation considering different tolls. Section 2.3 proposes three feedback control strategies based on the PI controller for iteratively adjusting the toll rates. Section 2.4 formulates the complete toll level problem and develops an SBO framework as the solution algorithm. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.